From 3ff92e5ca9d1aff4bf3c9f73f343ac5459e514a8 Mon Sep 17 00:00:00 2001 From: zhouxun Date: Tue, 14 Jul 2026 20:27:26 +0800 Subject: [PATCH] =?UTF-8?q?=E6=B7=BB=E5=8A=A0=2040=20=E7=AF=87=E8=AE=BA?= =?UTF-8?q?=E6=96=87=E7=A0=94=E7=A9=B6=E5=8D=A1=EF=BC=9AarXiv=20=E5=8D=81?= =?UTF-8?q?=E6=89=B9=E6=96=B0=E7=A0=94=E7=A9=B6?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Co-Authored-By: Claude Opus 4 --- SESSION-HANDOFF.md | 25 + data/note-index.json | 1328 ++++++++++++++++- .../papers/big-bench-hard-2022.json | 55 + data/review-receipts/papers/bigbird-2020.json | 55 + data/review-receipts/papers/bloom-2022.json | 55 + .../papers/controlnet-2023.json | 55 + .../papers/dreambooth-2022.json | 55 + data/review-receipts/papers/gorilla-2023.json | 55 + data/review-receipts/papers/gsm8k-2021.json | 55 + .../papers/hugginggpt-2023.json | 55 + .../papers/inner-monologue-2022.json | 55 + .../papers/kaplan-scaling-laws-2020.json | 55 + data/review-receipts/papers/lamda-2022.json | 55 + .../papers/least-to-most-prompting-2022.json | 55 + .../papers/linformer-2020.json | 55 + data/review-receipts/papers/longnet-2023.json | 55 + data/review-receipts/papers/minerva-2022.json | 55 + .../papers/mistral-7b-2023.json | 55 + .../papers/mrkl-systems-2022.json | 55 + .../papers/natural-instructions-v2-2022.json | 55 + data/review-receipts/papers/opt-2022.json | 55 + .../papers/orca-explanation-tuning-2023.json | 55 + .../papers/p-tuning-v2-2021.json | 55 + .../papers/pal-code-reasoning-2022.json | 55 + data/review-receipts/papers/palm-2022.json | 55 + .../papers/plan-and-solve-prompting-2023.json | 55 + .../papers/prefix-tuning-2021.json | 55 + .../papers/program-of-thoughts-2022.json | 55 + .../papers/prompt-to-prompt-2022.json | 55 + .../papers/prompt-tuning-2021.json | 55 + data/review-receipts/papers/qlora-2023.json | 55 + data/review-receipts/papers/saycan-2022.json | 55 + .../papers/self-instruct-2022.json | 55 + .../papers/speculative-decoding-2022.json | 55 + .../star-self-taught-reasoner-2022.json | 55 + .../papers/textual-inversion-2022.json | 55 + data/review-receipts/papers/toolllm-2023.json | 55 + data/review-receipts/papers/toxigen-2022.json | 55 + .../papers/truthfulqa-2021.json | 55 + data/review-receipts/papers/ul2-2022.json | 55 + data/review-receipts/papers/webgpt-2021.json | 55 + .../review-receipts/papers/wizardlm-2023.json | 55 + src/content/docs/about.md | 6 +- .../topic-papers-agents-and-llm-systems-01.md | 40 +- .../topic-papers-agents-and-llm-systems-02.md | 18 + ...pers-generative-models-and-diffusion-01.md | 8 +- ...c-papers-nlp-foundations-and-scaling-01.md | 6 +- ...c-papers-nlp-foundations-and-scaling-02.md | 7 +- src/content/docs/career-plan.md | 4 +- src/content/docs/index.md | 2 +- src/content/docs/method.md | 6 +- src/content/docs/papers-atlas.md | 16 +- src/content/docs/papers-method.md | 2 +- src/content/docs/papers-queue.md | 6 +- .../docs/papers/big-bench-hard-2022.md | 90 ++ src/content/docs/papers/bigbird-2020.md | 90 ++ src/content/docs/papers/bloom-2022.md | 90 ++ src/content/docs/papers/controlnet-2023.md | 90 ++ src/content/docs/papers/dreambooth-2022.md | 90 ++ src/content/docs/papers/gorilla-2023.md | 90 ++ src/content/docs/papers/gsm8k-2021.md | 90 ++ src/content/docs/papers/hugginggpt-2023.md | 90 ++ .../docs/papers/inner-monologue-2022.md | 90 ++ .../docs/papers/kaplan-scaling-laws-2020.md | 90 ++ src/content/docs/papers/lamda-2022.md | 90 ++ .../papers/least-to-most-prompting-2022.md | 90 ++ src/content/docs/papers/linformer-2020.md | 90 ++ src/content/docs/papers/longnet-2023.md | 90 ++ src/content/docs/papers/minerva-2022.md | 90 ++ src/content/docs/papers/mistral-7b-2023.md | 90 ++ src/content/docs/papers/mrkl-systems-2022.md | 90 ++ .../papers/natural-instructions-v2-2022.md | 90 ++ src/content/docs/papers/opt-2022.md | 90 ++ .../papers/orca-explanation-tuning-2023.md | 90 ++ src/content/docs/papers/p-tuning-v2-2021.md | 90 ++ .../docs/papers/pal-code-reasoning-2022.md | 90 ++ src/content/docs/papers/palm-2022.md | 90 ++ .../papers/plan-and-solve-prompting-2023.md | 90 ++ src/content/docs/papers/prefix-tuning-2021.md | 90 ++ .../docs/papers/program-of-thoughts-2022.md | 90 ++ .../docs/papers/prompt-to-prompt-2022.md | 90 ++ src/content/docs/papers/prompt-tuning-2021.md | 90 ++ src/content/docs/papers/qlora-2023.md | 90 ++ src/content/docs/papers/saycan-2022.md | 90 ++ src/content/docs/papers/self-instruct-2022.md | 90 ++ .../docs/papers/speculative-decoding-2022.md | 90 ++ .../papers/star-self-taught-reasoner-2022.md | 90 ++ .../docs/papers/textual-inversion-2022.md | 90 ++ src/content/docs/papers/toolllm-2023.md | 90 ++ src/content/docs/papers/toxigen-2022.md | 90 ++ src/content/docs/papers/truthfulqa-2021.md | 90 ++ src/content/docs/papers/ul2-2022.md | 90 ++ src/content/docs/papers/webgpt-2021.md | 90 ++ src/content/docs/papers/wizardlm-2023.md | 90 ++ src/content/docs/queue.md | 4 +- 95 files changed, 7228 insertions(+), 50 deletions(-) create mode 100644 data/review-receipts/papers/big-bench-hard-2022.json create mode 100644 data/review-receipts/papers/bigbird-2020.json create mode 100644 data/review-receipts/papers/bloom-2022.json create mode 100644 data/review-receipts/papers/controlnet-2023.json create mode 100644 data/review-receipts/papers/dreambooth-2022.json create mode 100644 data/review-receipts/papers/gorilla-2023.json create mode 100644 data/review-receipts/papers/gsm8k-2021.json create mode 100644 data/review-receipts/papers/hugginggpt-2023.json create mode 100644 data/review-receipts/papers/inner-monologue-2022.json create mode 100644 data/review-receipts/papers/kaplan-scaling-laws-2020.json create mode 100644 data/review-receipts/papers/lamda-2022.json create mode 100644 data/review-receipts/papers/least-to-most-prompting-2022.json create mode 100644 data/review-receipts/papers/linformer-2020.json create mode 100644 data/review-receipts/papers/longnet-2023.json create mode 100644 data/review-receipts/papers/minerva-2022.json create mode 100644 data/review-receipts/papers/mistral-7b-2023.json create mode 100644 data/review-receipts/papers/mrkl-systems-2022.json create mode 100644 data/review-receipts/papers/natural-instructions-v2-2022.json create mode 100644 data/review-receipts/papers/opt-2022.json create mode 100644 data/review-receipts/papers/orca-explanation-tuning-2023.json create mode 100644 data/review-receipts/papers/p-tuning-v2-2021.json create mode 100644 data/review-receipts/papers/pal-code-reasoning-2022.json create mode 100644 data/review-receipts/papers/palm-2022.json create mode 100644 data/review-receipts/papers/plan-and-solve-prompting-2023.json create mode 100644 data/review-receipts/papers/prefix-tuning-2021.json create mode 100644 data/review-receipts/papers/program-of-thoughts-2022.json create mode 100644 data/review-receipts/papers/prompt-to-prompt-2022.json create mode 100644 data/review-receipts/papers/prompt-tuning-2021.json create mode 100644 data/review-receipts/papers/qlora-2023.json create mode 100644 data/review-receipts/papers/saycan-2022.json create mode 100644 data/review-receipts/papers/self-instruct-2022.json create mode 100644 data/review-receipts/papers/speculative-decoding-2022.json create mode 100644 data/review-receipts/papers/star-self-taught-reasoner-2022.json create mode 100644 data/review-receipts/papers/textual-inversion-2022.json create mode 100644 data/review-receipts/papers/toolllm-2023.json create mode 100644 data/review-receipts/papers/toxigen-2022.json create mode 100644 data/review-receipts/papers/truthfulqa-2021.json create mode 100644 data/review-receipts/papers/ul2-2022.json create mode 100644 data/review-receipts/papers/webgpt-2021.json create mode 100644 data/review-receipts/papers/wizardlm-2023.json create mode 100644 src/content/docs/atlas/papers/topic-papers-agents-and-llm-systems-02.md create mode 100644 src/content/docs/papers/big-bench-hard-2022.md create mode 100644 src/content/docs/papers/bigbird-2020.md create mode 100644 src/content/docs/papers/bloom-2022.md create mode 100644 src/content/docs/papers/controlnet-2023.md create mode 100644 src/content/docs/papers/dreambooth-2022.md create mode 100644 src/content/docs/papers/gorilla-2023.md create mode 100644 src/content/docs/papers/gsm8k-2021.md create mode 100644 src/content/docs/papers/hugginggpt-2023.md create mode 100644 src/content/docs/papers/inner-monologue-2022.md create mode 100644 src/content/docs/papers/kaplan-scaling-laws-2020.md create mode 100644 src/content/docs/papers/lamda-2022.md create mode 100644 src/content/docs/papers/least-to-most-prompting-2022.md create mode 100644 src/content/docs/papers/linformer-2020.md create mode 100644 src/content/docs/papers/longnet-2023.md create mode 100644 src/content/docs/papers/minerva-2022.md create mode 100644 src/content/docs/papers/mistral-7b-2023.md create mode 100644 src/content/docs/papers/mrkl-systems-2022.md create mode 100644 src/content/docs/papers/natural-instructions-v2-2022.md create mode 100644 src/content/docs/papers/opt-2022.md create mode 100644 src/content/docs/papers/orca-explanation-tuning-2023.md create mode 100644 src/content/docs/papers/p-tuning-v2-2021.md create mode 100644 src/content/docs/papers/pal-code-reasoning-2022.md create mode 100644 src/content/docs/papers/palm-2022.md create mode 100644 src/content/docs/papers/plan-and-solve-prompting-2023.md create mode 100644 src/content/docs/papers/prefix-tuning-2021.md create mode 100644 src/content/docs/papers/program-of-thoughts-2022.md create mode 100644 src/content/docs/papers/prompt-to-prompt-2022.md create mode 100644 src/content/docs/papers/prompt-tuning-2021.md create mode 100644 src/content/docs/papers/qlora-2023.md create mode 100644 src/content/docs/papers/saycan-2022.md create mode 100644 src/content/docs/papers/self-instruct-2022.md create mode 100644 src/content/docs/papers/speculative-decoding-2022.md create mode 100644 src/content/docs/papers/star-self-taught-reasoner-2022.md create mode 100644 src/content/docs/papers/textual-inversion-2022.md create mode 100644 src/content/docs/papers/toolllm-2023.md create mode 100644 src/content/docs/papers/toxigen-2022.md create mode 100644 src/content/docs/papers/truthfulqa-2021.md create mode 100644 src/content/docs/papers/ul2-2022.md create mode 100644 src/content/docs/papers/webgpt-2021.md create mode 100644 src/content/docs/papers/wizardlm-2023.md diff --git a/SESSION-HANDOFF.md b/SESSION-HANDOFF.md index 5c0c4e20b..f11775ce8 100644 --- a/SESSION-HANDOFF.md +++ b/SESSION-HANDOFF.md @@ -2,6 +2,31 @@ > 状态：当前接班入口。旧的批量生产 session 快照已失效，不得用于恢复自动循环；持续运行使用只读 supervisor + 有界 writer epoch。 +## 2026-07-14 新增 40 篇论文与部署 Epoch Contract + +- status：`running` +- objective：在用户明确授权“分十批新研究 40 篇论文，全流程部署”下，新增 10 批 × 4 篇公开 arXiv 论文研究卡，覆盖 foundation/scaling、开放模型、instruction tuning、reasoning prompt、agent/tool use、PEFT、长上下文/推理、多模态生成与评测安全。 +- scope：允许新增 `src/content/docs/papers/*.md`、`data/review-receipts/papers/*.json`，刷新 `data/note-index.json` 与 papers atlas 派生页，同步公开规模文案和本 handoff；不修改候选队列、policy/threshold、既有论文正文语义或远端配置。 +- activated_by：`explicit-user-request-2026-07-14-new-40-papers-full-deploy` +- review_after：`2026-07-14` +- acceptance_checks： + - arXiv API 元数据校验：40/40 条目可解析； + - `node scripts/quality-gate.mjs --changed-from main --json`：checked=40, pass=true； + - 40 份 `study-review-receipt-v1` 的 canonical note digest 与正文一致； + - `npm run atlas`：2024 notes, 69 chunks； + - `npm run audit:counts`； + - `npm run audit:content-contract`； + - `npm run audit:links`； + - `npm run audit:wikilinks`； + - `git ls-files -co --exclude-standard -z | node scripts/audit-public-redlines.mjs --stdin0`； + - `npm run build:strict -- --log /tmp/study-forty-build-clean.log`； + - `git diff --check`； + - 提交后使用 `STUDY_CHANGED_FROM=384787e09827c336baf5ac2b33e67e8c91b9df49 npm run verify:ci` 做 PR/Pages portable gate。 +- budget：10 个内容小批次、40 篇新增 paper、1 个可写切片、1 个本地 writer、1 次 branch/PR/merge/deploy 窗口。 +- external_outcome：40 篇新增论文笔记进入公开 study 站点；验证状态保持 `UNVERIFIED`，不声明实际运行论文 benchmark。 +- stop_conditions：规范 Node/npm 不可用；arXiv 来源不可核验；内容契约、红线审计、strict build 或 verify:ci 失败且无法在本 scope 内修复；需要修改 policy/threshold、候选队列或隐私敏感内容；远端 CI/Pages 失败且需要新权限；用户停止。 +- superseded_by：`none` + ## 2026-07-14 新增 4 篇论文与部署 Epoch Contract - status：`running` diff --git a/data/note-index.json b/data/note-index.json index 92d87edab..e543081ee 100644 --- a/data/note-index.json +++ b/data/note-index.json @@ -3,16 +3,16 @@ "taxonomy_version": "taxonomy-v1", "stats": { "summary": { - "total": 1984, - "classified": 1937, + "total": 2024, + "classified": 1977, "unclassified": 47, "unknown_difficulty": 1975, "empty_description": 1970 }, "by_area": { "papers": { - "total": 1023, - "classified": 1004, + "total": 1063, + "classified": 1044, "unclassified": 19, "unknown_difficulty": 1014, "empty_description": 1013 @@ -27,7 +27,7 @@ }, "missing_curated_assignments": [], "atlas": { - "chunks": 68, + "chunks": 69, "chunk_size": 100, "max_chunk_entries": 100 } @@ -2596,6 +2596,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-formal-methods-01/" } }, + { + "id": "papers::big-bench-hard-2022", + "area": "papers", + "slug": "big-bench-hard-2022", + "title": "BIG-Bench Hard — 从大题库里挑出模型最头疼的 23 类题", + "description": "用 BBH 理解为什么 benchmark 需要难题子集和 CoT 对照。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Evaluation" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/big-bench-hard-2022/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::big-little-2011", "area": "papers", @@ -2660,6 +2692,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-hci-and-software-engineering-research-01/" } }, + { + "id": "papers::bigbird-2020", + "area": "papers", + "slug": "bigbird-2020", + "title": "BigBird — 用稀疏 attention 拉长 Transformer 视野", + "description": "用 BigBird 理解局部、全局和随机 attention 怎样组成长序列模式。", + "difficulty": "advanced", + "canonical_topics": [ + "papers-nlp-foundations-and-scaling" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-nlp-foundations-and-scaling", + "matched_category": "nlp", + "raw_category": "NLP / Efficient Attention" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/bigbird-2020/", + "atlas": { + "chunk_id": "topic-papers-nlp-foundations-and-scaling-01", + "chunk_route": "/study/atlas/papers/topic-papers-nlp-foundations-and-scaling-01/" + } + }, { "id": "papers::biggan-2018", "area": "papers", @@ -3010,6 +3074,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-information-retrieval-and-recommendation-01/" } }, + { + "id": "papers::bloom-2022", + "area": "papers", + "slug": "bloom-2022", + "title": "BLOOM — 把 176B 多语种模型做成开放科学工程", + "description": "用 BLOOM 理解大模型也可以用社区协作、数据治理和开放发布来推进。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Open Science" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/bloom-2022/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::bm25", "area": "papers", @@ -6226,6 +6322,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-cryptography-and-security-01/" } }, + { + "id": "papers::controlnet-2023", + "area": "papers", + "slug": "controlnet-2023", + "title": "ControlNet — 给扩散模型加一条可控条件支路", + "description": "用 ControlNet 理解边缘、姿态和深度图如何稳定控制图像生成。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-generative-models-and-diffusion" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-generative-models-and-diffusion", + "matched_category": "diffusion", + "raw_category": "Diffusion / Control" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/controlnet-2023/", + "atlas": { + "chunk_id": "topic-papers-generative-models-and-diffusion-01", + "chunk_route": "/study/atlas/papers/topic-papers-generative-models-and-diffusion-01/" + } + }, { "id": "papers::cook-1984-distributed-ray-tracing", "area": "papers", @@ -8993,6 +9121,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-reinforcement-learning-01/" } }, + { + "id": "papers::dreambooth-2022", + "area": "papers", + "slug": "dreambooth-2022", + "title": "DreamBooth — 用几张图把一个新主体塞进生成模型", + "description": "用 DreamBooth 理解 subject-driven generation 怎样让扩散模型记住特定对象。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-generative-models-and-diffusion" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-generative-models-and-diffusion", + "matched_category": "diffusion", + "raw_category": "Diffusion / Personalization" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/dreambooth-2022/", + "atlas": { + "chunk_id": "topic-papers-generative-models-and-diffusion-01", + "chunk_route": "/study/atlas/papers/topic-papers-generative-models-and-diffusion-01/" + } + }, { "id": "papers::dreamfusion-2022", "area": "papers", @@ -12660,6 +12820,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-computer-graphics-and-visualization-01/" } }, + { + "id": "papers::gorilla-2023", + "area": "papers", + "slug": "gorilla-2023", + "title": "Gorilla — 让 LLM 学会查 API 文档再调用", + "description": "用 Gorilla 理解 API grounding 如何降低工具调用幻觉。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Tool Use" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/gorilla-2023/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::gortler-1996-lumigraph", "area": "papers", @@ -13326,6 +13518,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-nlp-foundations-and-scaling-01/" } }, + { + "id": "papers::gsm8k-2021", + "area": "papers", + "slug": "gsm8k-2021", + "title": "GSM8K — 小学数学题把大模型算术短板照出来", + "description": "用 GSM8K 理解数学 word problem、verifier 和采样重排为什么重要。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Math Evaluation" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/gsm8k-2021/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::hacl-star-2017", "area": "papers", @@ -14342,6 +14566,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-information-and-coding-theory-01/" } }, + { + "id": "papers::hugginggpt-2023", + "area": "papers", + "slug": "hugginggpt-2023", + "title": "HuggingGPT — 让 ChatGPT 当任务调度员，模型库当工具箱", + "description": "用 HuggingGPT 理解 LLM 如何规划并调用专用模型完成多模态任务。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Tool Orchestration" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/hugginggpt-2023/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::hughes-fp-matters", "area": "papers", @@ -14754,6 +15010,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-databases-01/" } }, + { + "id": "papers::inner-monologue-2022", + "area": "papers", + "slug": "inner-monologue-2022", + "title": "Inner Monologue — 让机器人把观察结果说回计划里", + "description": "用 Inner Monologue 理解闭环反馈如何让语言计划接上真实环境变化。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "agent", + "raw_category": "LLM Agent / Robotics" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/inner-monologue-2022/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::instant-ngp-2022", "area": "papers", @@ -15516,6 +15804,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-compilers-and-programming-language-theory-01/" } }, + { + "id": "papers::kaplan-scaling-laws-2020", + "area": "papers", + "slug": "kaplan-scaling-laws-2020", + "title": "Scaling Laws — 大模型训练不是玄学，是幂律预算题", + "description": "用 Kaplan scaling laws 理解参数、数据和计算量怎样一起决定语言模型损失。", + "difficulty": "advanced", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Scaling Laws" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/kaplan-scaling-laws-2020/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::karger-1997-consistent-hashing", "area": "papers", @@ -16467,6 +16787,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-information-retrieval-and-recommendation-01/" } }, + { + "id": "papers::lamda-2022", + "area": "papers", + "slug": "lamda-2022", + "title": "LaMDA — 聊天模型先学会有用、具体和不乱说", + "description": "用 LaMDA 理解开放域对话模型为什么需要质量、安全和 groundedness 三条线。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Dialogue" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/lamda-2022/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::lamport-1978", "area": "papers", @@ -16785,6 +17137,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-compilers-and-programming-language-theory-01/" } }, + { + "id": "papers::least-to-most-prompting-2022", + "area": "papers", + "slug": "least-to-most-prompting-2022", + "title": "Least-to-Most — 先拆小题，再解大题", + "description": "用 Least-to-Most Prompting 理解复杂推理为什么要先分解再逐步求解。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Reasoning" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/least-to-most-prompting-2022/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::lee-keystone-2020", "area": "papers", @@ -17200,6 +17584,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-distributed-systems-01/" } }, + { + "id": "papers::linformer-2020", + "area": "papers", + "slug": "linformer-2020", + "title": "Linformer — 把 attention 矩阵投影成线性复杂度", + "description": "用 Linformer 理解低秩假设如何压缩 self-attention。", + "difficulty": "advanced", + "canonical_topics": [ + "papers-nlp-foundations-and-scaling" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-nlp-foundations-and-scaling", + "matched_category": "nlp", + "raw_category": "NLP / Efficient Attention" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/linformer-2020/", + "atlas": { + "chunk_id": "topic-papers-nlp-foundations-and-scaling-01", + "chunk_route": "/study/atlas/papers/topic-papers-nlp-foundations-and-scaling-01/" + } + }, { "id": "papers::linux-kernel", "area": "papers", @@ -17802,6 +18218,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-nlp-foundations-and-scaling-01/" } }, + { + "id": "papers::longnet-2023", + "area": "papers", + "slug": "longnet-2023", + "title": "LongNet — 用 dilated attention 把上下文推到十亿 token 想象空间", + "description": "用 LongNet 理解扩张式 attention 如何在多尺度上连接超长序列。", + "difficulty": "advanced", + "canonical_topics": [ + "papers-nlp-foundations-and-scaling" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-nlp-foundations-and-scaling", + "matched_category": "nlp", + "raw_category": "NLP / Long Context" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/longnet-2023/", + "atlas": { + "chunk_id": "topic-papers-nlp-foundations-and-scaling-01", + "chunk_route": "/study/atlas/papers/topic-papers-nlp-foundations-and-scaling-01/" + } + }, { "id": "papers::loong-doc-mt", "area": "papers", @@ -19780,6 +20228,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-compilers-and-programming-language-theory-01/" } }, + { + "id": "papers::minerva-2022", + "area": "papers", + "slug": "minerva-2022", + "title": "Minerva — 把语言模型拉进数学草稿纸", + "description": "用 Minerva 理解为什么数学推理需要专门的数据、逐步解题和采样验证。", + "difficulty": "advanced", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Math Reasoning" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/minerva-2022/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::minhash-broder-1997", "area": "papers", @@ -20004,6 +20484,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" } }, + { + "id": "papers::mistral-7b-2023", + "area": "papers", + "slug": "mistral-7b-2023", + "title": "Mistral 7B — 小模型靠架构细节打出性价比", + "description": "用 Mistral 7B 理解 grouped-query attention 和 sliding-window attention 如何服务高效开源模型。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Efficient Model" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/mistral-7b-2023/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::mitls-2014-triple-handshake", "area": "papers", @@ -20547,6 +21059,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-network-protocols-01/" } }, + { + "id": "papers::mrkl-systems-2022", + "area": "papers", + "slug": "mrkl-systems-2022", + "title": "MRKL — 给大模型配一组专家工具和路由器", + "description": "用 MRKL Systems 理解 neuro-symbolic agent 为什么要把 LLM、检索和计算模块拆开。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Tool Architecture" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/mrkl-systems-2022/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::ms-marco-2016", "area": "papers", @@ -20898,6 +21442,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-distributed-systems-01/" } }, + { + "id": "papers::natural-instructions-v2-2022", + "area": "papers", + "slug": "natural-instructions-v2-2022", + "title": "Super-NaturalInstructions — 1600+ 任务教模型读懂说明书", + "description": "用 Super-NaturalInstructions 理解 declarative instructions 如何评测任务泛化。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Instruction Benchmark" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/natural-instructions-v2-2022/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::nbeats-2020", "area": "papers", @@ -21978,6 +22554,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-operating-systems-and-cluster-management-01/" } }, + { + "id": "papers::opt-2022", + "area": "papers", + "slug": "opt-2022", + "title": "OPT — 把 GPT-3 级训练日志打开给研究社区", + "description": "用 OPT 理解开放权重、训练日志和复现实验对 LLM 研究的重要性。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Open Model" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/opt-2022/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::optuna", "area": "papers", @@ -22074,6 +22682,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" } }, + { + "id": "papers::orca-explanation-tuning-2023", + "area": "papers", + "slug": "orca-explanation-tuning-2023", + "title": "Orca — 小模型不只抄答案，还学解释轨迹", + "description": "用 Orca 理解 explanation tuning 为什么比只蒸馏最终答案更像教学生。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Distillation" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/orca-explanation-tuning-2023/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::oscar-int2-kv", "area": "papers", @@ -22202,6 +22842,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-distributed-training-and-gpu-01/" } }, + { + "id": "papers::p-tuning-v2-2021", + "area": "papers", + "slug": "p-tuning-v2-2021", + "title": "P-Tuning v2 — 把 prompt tuning 深插到每一层", + "description": "用 P-Tuning v2 理解深层连续提示为什么能跨规模和任务接近 fine-tuning。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Efficient Finetuning" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/p-tuning-v2-2021/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::p4-2014", "area": "papers", @@ -22425,6 +23097,70 @@ "chunk_route": "/study/atlas/papers/topic-papers-hci-and-software-engineering-research-01/" } }, + { + "id": "papers::pal-code-reasoning-2022", + "area": "papers", + "slug": "pal-code-reasoning-2022", + "title": "PAL — 让 Python 成为语言模型的草稿纸", + "description": "用 PAL 理解 Program-aided Language Models 如何把推理转成可运行代码。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Tool Reasoning" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/pal-code-reasoning-2022/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, + { + "id": "papers::palm-2022", + "area": "papers", + "slug": "palm-2022", + "title": "PaLM — Pathways 把 540B LLM 扩成统一底座", + "description": "用 PaLM 理解大规模 dense decoder 如何在多任务、推理和代码能力上同时冒头。", + "difficulty": "advanced", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Foundation Model" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/palm-2022/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::panel", "area": "papers", @@ -23192,6 +23928,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-distributed-systems-01/" } }, + { + "id": "papers::plan-and-solve-prompting-2023", + "area": "papers", + "slug": "plan-and-solve-prompting-2023", + "title": "Plan-and-Solve — 零样本推理先写计划再执行", + "description": "用 Plan-and-Solve 理解为什么 prompt 可以显式拆成 plan 和 solve 两段。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Reasoning" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/plan-and-solve-prompting-2023/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::plan9-1995", "area": "papers", @@ -23539,6 +24307,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-nlp-foundations-and-scaling-01/" } }, + { + "id": "papers::prefix-tuning-2021", + "area": "papers", + "slug": "prefix-tuning-2021", + "title": "Prefix-Tuning — 不改模型，只给每层塞一段可训练前缀", + "description": "用 Prefix-Tuning 理解连续 prompt 如何成为参数高效微调方法。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Efficient Finetuning" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/prefix-tuning-2021/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::presumed-abort-1986", "area": "papers", @@ -23634,6 +24434,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-hci-and-software-engineering-research-01/" } }, + { + "id": "papers::program-of-thoughts-2022", + "area": "papers", + "slug": "program-of-thoughts-2022", + "title": "Program of Thoughts — 让模型写程序，把计算交给解释器", + "description": "用 Program of Thoughts 理解自然语言推理和精确计算为什么要分工。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Tool Reasoning" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/program-of-thoughts-2022/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::program-shepherding-2002", "area": "papers", @@ -23761,6 +24593,70 @@ "chunk_route": "/study/atlas/papers/topic-papers-compilers-and-programming-language-theory-01/" } }, + { + "id": "papers::prompt-to-prompt-2022", + "area": "papers", + "slug": "prompt-to-prompt-2022", + "title": "Prompt-to-Prompt — 改词不改构图的 cross-attention 编辑", + "description": "用 Prompt-to-Prompt 理解扩散模型里文本 token 和图像布局如何对齐。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-generative-models-and-diffusion" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-generative-models-and-diffusion", + "matched_category": "diffusion", + "raw_category": "Diffusion / Editing" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/prompt-to-prompt-2022/", + "atlas": { + "chunk_id": "topic-papers-generative-models-and-diffusion-01", + "chunk_route": "/study/atlas/papers/topic-papers-generative-models-and-diffusion-01/" + } + }, + { + "id": "papers::prompt-tuning-2021", + "area": "papers", + "slug": "prompt-tuning-2021", + "title": "Prompt Tuning — 规模变大后，软提示也能接近微调", + "description": "用 Prompt Tuning 理解为什么 soft prompt 在大模型上突然变得有效。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Efficient Finetuning" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/prompt-tuning-2021/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::prototypical-networks-2017", "area": "papers", @@ -23920,6 +24816,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-compilers-and-programming-language-theory-01/" } }, + { + "id": "papers::qlora-2023", + "area": "papers", + "slug": "qlora-2023", + "title": "QLoRA — 4-bit 量化底座上贴 LoRA 也能微调", + "description": "用 QLoRA 理解 NF4、double quantization 和 paged optimizers 如何降低微调门槛。", + "difficulty": "advanced", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Efficient Finetuning" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/qlora-2023/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::quantum-supremacy-2019", "area": "papers", @@ -25919,6 +26847,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-information-retrieval-and-recommendation-01/" } }, + { + "id": "papers::saycan-2022", + "area": "papers", + "slug": "saycan-2022", + "title": "SayCan — 机器人不只问“想做什么”，还问“我能做什么”", + "description": "用 SayCan 理解语言模型和机器人 affordance 如何合成可执行动作。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "agent", + "raw_category": "LLM Agent / Robotics" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/saycan-2022/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::scads-database-2008", "area": "papers", @@ -26461,6 +27421,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" } }, + { + "id": "papers::self-instruct-2022", + "area": "papers", + "slug": "self-instruct-2022", + "title": "Self-Instruct — 让模型自己造指令数据再学习", + "description": "用 Self-Instruct 理解指令微调数据如何从少量种子任务扩展出来。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Instruction Tuning" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/self-instruct-2022/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::self-pic", "area": "papers", @@ -28208,6 +29200,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" } }, + { + "id": "papers::speculative-decoding-2022", + "area": "papers", + "slug": "speculative-decoding-2022", + "title": "Speculative Decoding — 小模型先猜，大模型只验收", + "description": "用 Speculative Decoding 理解如何不改变分布地加速自回归生成。", + "difficulty": "advanced", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Inference" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/speculative-decoding-2022/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::splade-2021", "area": "papers", @@ -28461,6 +29485,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-compilers-and-programming-language-theory-02/" } }, + { + "id": "papers::star-self-taught-reasoner-2022", + "area": "papers", + "slug": "star-self-taught-reasoner-2022", + "title": "STaR — 模型先试着讲理由，再用对的理由训练自己", + "description": "用 STaR 理解 rationale bootstrapping 怎样减少人工推理标注。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Reasoning" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/star-self-taught-reasoner-2022/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::starcoder-2023", "area": "papers", @@ -28967,8 +30023,8 @@ }, "route": "/study/papers/sycophancy-2023/", "atlas": { - "chunk_id": "topic-papers-nlp-foundations-and-scaling-01", - "chunk_route": "/study/atlas/papers/topic-papers-nlp-foundations-and-scaling-01/" + "chunk_id": "topic-papers-nlp-foundations-and-scaling-02", + "chunk_route": "/study/atlas/papers/topic-papers-nlp-foundations-and-scaling-02/" } }, { @@ -29127,8 +30183,8 @@ }, "route": "/study/papers/t0-2021/", "atlas": { - "chunk_id": "topic-papers-nlp-foundations-and-scaling-01", - "chunk_route": "/study/atlas/papers/topic-papers-nlp-foundations-and-scaling-01/" + "chunk_id": "topic-papers-nlp-foundations-and-scaling-02", + "chunk_route": "/study/atlas/papers/topic-papers-nlp-foundations-and-scaling-02/" } }, { @@ -29158,8 +30214,8 @@ }, "route": "/study/papers/t5/", "atlas": { - "chunk_id": "topic-papers-nlp-foundations-and-scaling-01", - "chunk_route": "/study/atlas/papers/topic-papers-nlp-foundations-and-scaling-01/" + "chunk_id": "topic-papers-nlp-foundations-and-scaling-02", + "chunk_route": "/study/atlas/papers/topic-papers-nlp-foundations-and-scaling-02/" } }, { @@ -29609,6 +30665,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-operating-systems-and-cluster-management-01/" } }, + { + "id": "papers::textual-inversion-2022", + "area": "papers", + "slug": "textual-inversion-2022", + "title": "Textual Inversion — 给新概念学一个专属 token", + "description": "用 Textual Inversion 理解冻结扩散模型时如何只学习概念 embedding。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-generative-models-and-diffusion" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-generative-models-and-diffusion", + "matched_category": "diffusion", + "raw_category": "Diffusion / Personalization" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/textual-inversion-2022/", + "atlas": { + "chunk_id": "topic-papers-generative-models-and-diffusion-01", + "chunk_route": "/study/atlas/papers/topic-papers-generative-models-and-diffusion-01/" + } + }, { "id": "papers::tfidf-classic", "area": "papers", @@ -30084,6 +31172,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" } }, + { + "id": "papers::toolllm-2023", + "area": "papers", + "slug": "toolllm-2023", + "title": "ToolLLM — 用 16000+ API 训练模型进入真实工具世界", + "description": "用 ToolLLM 理解大规模 API 数据集、工具检索和工具评测如何支撑 agent。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Tool Use" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/toolllm-2023/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::tor-2004", "area": "papers", @@ -30116,6 +31236,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-network-protocols-01/" } }, + { + "id": "papers::toxigen-2022", + "area": "papers", + "slug": "toxigen-2022", + "title": "ToxiGen — 用生成模型造隐性仇恨测试集", + "description": "用 ToxiGen 理解安全评测为什么要覆盖隐性、对抗性和群体相关文本。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Safety Evaluation" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/toxigen-2022/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::toy-models-superposition", "area": "papers", @@ -30530,6 +31682,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-information-retrieval-and-recommendation-01/" } }, + { + "id": "papers::truthfulqa-2021", + "area": "papers", + "slug": "truthfulqa-2021", + "title": "TruthfulQA — 专门问模型容易学人类谬误的问题", + "description": "用 TruthfulQA 理解语言模型为什么会模仿常见假话而不是坚持事实。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Evaluation" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/truthfulqa-2021/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::turchin-supercompilation", "area": "papers", @@ -30753,6 +31937,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-databases-01/" } }, + { + "id": "papers::ul2-2022", + "area": "papers", + "slug": "ul2-2022", + "title": "UL2 — 一个模型同时练完补空、续写和长文本", + "description": "用 UL2 理解 mixture-of-denoisers 如何统一不同语言模型训练范式。", + "difficulty": "advanced", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Pretraining Objective" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/ul2-2022/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::unified-memory-2014", "area": "papers", @@ -31868,6 +33084,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-databases-01/" } }, + { + "id": "papers::webgpt-2021", + "area": "papers", + "slug": "webgpt-2021", + "title": "WebGPT — 让模型带着浏览器回答问题", + "description": "用 WebGPT 理解检索、引用和人类偏好如何组合成可追溯问答。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "agent", + "raw_category": "LLM / Browser Agent" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/webgpt-2021/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::websocket-rfc-6455", "area": "papers", @@ -32124,6 +33372,38 @@ "chunk_route": "/study/atlas/papers/topic-papers-network-protocols-01/" } }, + { + "id": "papers::wizardlm-2023", + "area": "papers", + "slug": "wizardlm-2023", + "title": "WizardLM — 用 Evol-Instruct 自动变难训练题", + "description": "用 WizardLM 理解 instruction 数据不只要多，还要逐步变复杂。", + "difficulty": "intermediate", + "canonical_topics": [ + "papers-agents-and-llm-systems" + ], + "classification": { + "state": "classified", + "source": "frontmatter-category", + "topic_id": "papers-agents-and-llm-systems", + "matched_category": "llm", + "raw_category": "LLM / Instruction Tuning" + }, + "trust": { + "contract_state": "v2", + "verification_status": "UNVERIFIED" + }, + "freshness": { + "state": "NOT_EVALUATED", + "reviewed_at": "2026-07-14", + "review_after": null + }, + "route": "/study/papers/wizardlm-2023/", + "atlas": { + "chunk_id": "topic-papers-agents-and-llm-systems-01", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + } + }, { "id": "papers::word2vec", "area": "papers", @@ -32565,8 +33845,8 @@ }, "route": "/study/papers/zombie-agents-2602/", "atlas": { - "chunk_id": "topic-papers-agents-and-llm-systems-01", - "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/" + "chunk_id": "topic-papers-agents-and-llm-systems-02", + "chunk_route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-02/" } }, { @@ -63035,9 +64315,23 @@ "en": "Agents and LLM Systems" }, "page": 1, - "pages": 1, + "pages": 2, "route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-01/", - "entries": 68 + "entries": 100 + }, + { + "id": "topic-papers-agents-and-llm-systems-02", + "area": "papers", + "kind": "topic", + "topic_id": "papers-agents-and-llm-systems", + "labels": { + "zh": "智能体与 LLM 系统", + "en": "Agents and LLM Systems" + }, + "page": 2, + "pages": 2, + "route": "/study/atlas/papers/topic-papers-agents-and-llm-systems-02/", + "entries": 1 }, { "id": "topic-papers-ai-safety-and-interpretability-01", @@ -63247,7 +64541,7 @@ "page": 1, "pages": 1, "route": "/study/atlas/papers/topic-papers-generative-models-and-diffusion-01/", - "entries": 10 + "entries": 14 }, { "id": "topic-papers-hci-and-software-engineering-research-01", @@ -63345,7 +64639,7 @@ "page": 2, "pages": 2, "route": "/study/atlas/papers/topic-papers-nlp-foundations-and-scaling-02/", - "entries": 11 + "entries": 14 }, { "id": "topic-papers-operating-systems-and-cluster-management-01", diff --git a/data/review-receipts/papers/big-bench-hard-2022.json b/data/review-receipts/papers/big-bench-hard-2022.json new file mode 100644 index 000000000..78ff52a56 --- /dev/null +++ b/data/review-receipts/papers/big-bench-hard-2022.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "big-bench-hard-2022", + "digest_sha256": "6d7d582826d9fa135f2d1c8e8603b6be560a6aa0f5f53558f6203f902d4ab2c6" + }, + "source_revision": "arXiv:2210.09261v1", + "research_input_sha256": "66ebd2a100d7b0beda16f53dc8dd7401edc2f2874ea1cb2011c136a517e555aa", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/bigbird-2020.json b/data/review-receipts/papers/bigbird-2020.json new file mode 100644 index 000000000..6d923eae8 --- /dev/null +++ b/data/review-receipts/papers/bigbird-2020.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "bigbird-2020", + "digest_sha256": "995a4d69e73eb9f098d8f9ebf5982bccec613f573acf9ebf597b41f6409699a8" + }, + "source_revision": "arXiv:2007.14062v2", + "research_input_sha256": "d744b073ba50ddce66b6e067b911ce4433c301048d15f1e47263de305b6ec012", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/bloom-2022.json b/data/review-receipts/papers/bloom-2022.json new file mode 100644 index 000000000..32538f1c8 --- /dev/null +++ b/data/review-receipts/papers/bloom-2022.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "bloom-2022", + "digest_sha256": "74aa5fefd63016af4b0cc709028a3118e00cb4e7daebf1a1064435392e1d18b7" + }, + "source_revision": "arXiv:2211.05100v4", + "research_input_sha256": "2d5b18618bc3f4dae1cc7814442b3da7f8c445270a53f3d16cc9b5e53f7c83fe", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/controlnet-2023.json b/data/review-receipts/papers/controlnet-2023.json new file mode 100644 index 000000000..fcaa35287 --- /dev/null +++ b/data/review-receipts/papers/controlnet-2023.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "controlnet-2023", + "digest_sha256": "2c4dfbaf90df781a252aad8d6e31fbe4d20918a97a3aa204722ab1675d883113" + }, + "source_revision": "arXiv:2302.05543v3", + "research_input_sha256": "0dfae02d64e50c17d5d50a7fa6c9cd8b05954c101e33169776b6381761781a72", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/dreambooth-2022.json b/data/review-receipts/papers/dreambooth-2022.json new file mode 100644 index 000000000..2e6b096ed --- /dev/null +++ b/data/review-receipts/papers/dreambooth-2022.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "dreambooth-2022", + "digest_sha256": "8224e32165149dc618fec8b5efefd918fcd01f46a69e2aeb581588aa85818c0e" + }, + "source_revision": "arXiv:2208.12242v2", + "research_input_sha256": "26e088769607243ef0f4b5dad39c327c7c5ac4ec1c91e05a388b05d9580f8b85", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/gorilla-2023.json b/data/review-receipts/papers/gorilla-2023.json new file mode 100644 index 000000000..90d862d8f --- /dev/null +++ b/data/review-receipts/papers/gorilla-2023.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "gorilla-2023", + "digest_sha256": "6150edbbabffdca5905d6b32a0bd4eef0dac774a1d1e23cf716355596ff05784" + }, + "source_revision": "arXiv:2305.15334v1", + "research_input_sha256": "b7b92dacdb5b5524c4fecf784fa9560ee61a0d18378393631841c0f57179bbbe", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/gsm8k-2021.json b/data/review-receipts/papers/gsm8k-2021.json new file mode 100644 index 000000000..6757fb09c --- /dev/null +++ b/data/review-receipts/papers/gsm8k-2021.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "gsm8k-2021", + "digest_sha256": "ce0b4e02509476e0d2bfbdeaaf2b7f7684c44405ef374cc13bd84c2d7147e0e7" + }, + "source_revision": "arXiv:2110.14168v2", + "research_input_sha256": "efb9a56f0ff87949a517bdb03118dece407326f31c60e04cd3a98b020ae40785", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/hugginggpt-2023.json b/data/review-receipts/papers/hugginggpt-2023.json new file mode 100644 index 000000000..b4b995c93 --- /dev/null +++ b/data/review-receipts/papers/hugginggpt-2023.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "hugginggpt-2023", + "digest_sha256": "d276c3384e7eb0128aeddc0387bc466f2857bfe2bb24ff52569266282fc78a44" + }, + "source_revision": "arXiv:2303.17580v4", + "research_input_sha256": "5fc26a9fb2d037e638ee0ab76db7488d659293518cd60adeb9da9b36af9d1c82", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/inner-monologue-2022.json b/data/review-receipts/papers/inner-monologue-2022.json new file mode 100644 index 000000000..05380230c --- /dev/null +++ b/data/review-receipts/papers/inner-monologue-2022.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "inner-monologue-2022", + "digest_sha256": "97e57acf5affaab65285d5d80ae469b5601743a3adab2f423c85dae2110ec3cb" + }, + "source_revision": "arXiv:2207.05608v1", + "research_input_sha256": "44063aa526ff4eed240cfb43292146864657d674e00b2e923cdc6e2a2d021fdf", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/kaplan-scaling-laws-2020.json b/data/review-receipts/papers/kaplan-scaling-laws-2020.json new file mode 100644 index 000000000..ea2866b9d --- /dev/null +++ b/data/review-receipts/papers/kaplan-scaling-laws-2020.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "kaplan-scaling-laws-2020", + "digest_sha256": "290943ad677bacccf3b829e9f3e8771eff0a5342278c65276d93fe6c89c5729b" + }, + "source_revision": "arXiv:2001.08361v1", + "research_input_sha256": "65167231bcdfc0382012ea45cd0fe4f3c9bff8862d9f0812f9d9c5c56dcc9357", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/lamda-2022.json b/data/review-receipts/papers/lamda-2022.json new file mode 100644 index 000000000..30355b489 --- /dev/null +++ b/data/review-receipts/papers/lamda-2022.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "lamda-2022", + "digest_sha256": "f21fb0b5917772720acaa133954545fe5677a4721c4b2dc4a8b5e81f874e9e0b" + }, + "source_revision": "arXiv:2201.08239v3", + "research_input_sha256": "c0c685404880afee4e9e893d24915fa5ce1397e3f6a4c2ac29a4b909417bb240", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/least-to-most-prompting-2022.json b/data/review-receipts/papers/least-to-most-prompting-2022.json new file mode 100644 index 000000000..1805337db --- /dev/null +++ b/data/review-receipts/papers/least-to-most-prompting-2022.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "least-to-most-prompting-2022", + "digest_sha256": "5f6e7e61e72ac38a137e3a2dff081be8cf7276ef5e04bc81c6600ca12cd6be3c" + }, + "source_revision": "arXiv:2205.10625v3", + "research_input_sha256": "5b22c3a273bcb0b7b503443adc6e4eb6d5d8df3e692e88e20edbadbe9691174c", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/linformer-2020.json b/data/review-receipts/papers/linformer-2020.json new file mode 100644 index 000000000..1aa78f1d1 --- /dev/null +++ b/data/review-receipts/papers/linformer-2020.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "linformer-2020", + "digest_sha256": "de526de470d111145d859cc69bcb2110e9f0268f22cfbdef15233cec44d2f308" + }, + "source_revision": "arXiv:2006.04768v3", + "research_input_sha256": "3cdf8c4489a0e3edff6eef56803b19b2fca69d9111b77172fd0274bd225f8676", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/longnet-2023.json b/data/review-receipts/papers/longnet-2023.json new file mode 100644 index 000000000..07910abd6 --- /dev/null +++ b/data/review-receipts/papers/longnet-2023.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "longnet-2023", + "digest_sha256": "bb44d774d93dd128af704f0d2152294291d694822f19bc484f7456883a7db5ac" + }, + "source_revision": "arXiv:2307.02486v2", + "research_input_sha256": "6f18d710038430180361e1fd013e29864abe7f6a624452ff2bc9e7cf657770bd", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/minerva-2022.json b/data/review-receipts/papers/minerva-2022.json new file mode 100644 index 000000000..978d29d8c --- /dev/null +++ b/data/review-receipts/papers/minerva-2022.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "minerva-2022", + "digest_sha256": "dca9c2d973a768c36acff3bf33195b500832857e569d639e6884da74df34d906" + }, + "source_revision": "arXiv:2206.14858v2", + "research_input_sha256": "981905e4676af347d3e699e5df252f824ab8b3c1a015e6680f69279eb9285f00", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/mistral-7b-2023.json b/data/review-receipts/papers/mistral-7b-2023.json new file mode 100644 index 000000000..40b5e849a --- /dev/null +++ b/data/review-receipts/papers/mistral-7b-2023.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "mistral-7b-2023", + "digest_sha256": "7ca28dbf4f5fda56c45882bf5d68369a13ad0785fa5b2c21a6e8c4a0f714a13d" + }, + "source_revision": "arXiv:2310.06825v1", + "research_input_sha256": "8152a932b02ec3408b2e752a4869a4b3562167bb60576c41cdb5b0c5f6af1b0e", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/mrkl-systems-2022.json b/data/review-receipts/papers/mrkl-systems-2022.json new file mode 100644 index 000000000..1d18d8854 --- /dev/null +++ b/data/review-receipts/papers/mrkl-systems-2022.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "mrkl-systems-2022", + "digest_sha256": "0f275a7b87ac803d9219d9d4118c9e37dbe19b27c66cf43647f2d946a646f4c4" + }, + "source_revision": "arXiv:2205.00445v1", + "research_input_sha256": "ba86ab684e69bcf3d4795a93b4fd2048ef89bf1f6c30dea443977dcaa5ba2b16", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/natural-instructions-v2-2022.json b/data/review-receipts/papers/natural-instructions-v2-2022.json new file mode 100644 index 000000000..2407b3440 --- /dev/null +++ b/data/review-receipts/papers/natural-instructions-v2-2022.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "natural-instructions-v2-2022", + "digest_sha256": "b2c0b67d86aef722c0d89cf12804ba7e222b8274ad8a75ff202614d90148619b" + }, + "source_revision": "arXiv:2204.07705v3", + "research_input_sha256": "130079a0d10b62ae7cb8c4bc3089b1fee767aca2ce534913e42f705840c4682a", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/opt-2022.json b/data/review-receipts/papers/opt-2022.json new file mode 100644 index 000000000..f8ac4dfbb --- /dev/null +++ b/data/review-receipts/papers/opt-2022.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "opt-2022", + "digest_sha256": "869006ce0fbf9be2de2bf590cc2a737000569259ce246a332317352541ad1587" + }, + "source_revision": "arXiv:2205.01068v4", + "research_input_sha256": "473ea8798985546dd37d15e6246f29b91edab40a619aa45972664e415e80fa7c", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/orca-explanation-tuning-2023.json b/data/review-receipts/papers/orca-explanation-tuning-2023.json new file mode 100644 index 000000000..7469c0e22 --- /dev/null +++ b/data/review-receipts/papers/orca-explanation-tuning-2023.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "orca-explanation-tuning-2023", + "digest_sha256": "05b36217e949245ed2afe74ce10fc9a06141ff9f2e748e22c6f1ce319c33288b" + }, + "source_revision": "arXiv:2306.02707v1", + "research_input_sha256": "d10bcd6d1110c5034dea0d701d7355e90e5297641492c0bc064d9675b9fb2d6b", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/p-tuning-v2-2021.json b/data/review-receipts/papers/p-tuning-v2-2021.json new file mode 100644 index 000000000..735eccaad --- /dev/null +++ b/data/review-receipts/papers/p-tuning-v2-2021.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "p-tuning-v2-2021", + "digest_sha256": "7e29970dd1c54b08232dcacf0424d15144db65fddf56a9513e349bcc2d66fc61" + }, + "source_revision": "arXiv:2110.07602v3", + "research_input_sha256": "1cded8dc23b7fc4e5e2ca05fba142265ad7d5af2230f76394fea4e7598ab84a5", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/pal-code-reasoning-2022.json b/data/review-receipts/papers/pal-code-reasoning-2022.json new file mode 100644 index 000000000..30184aafd --- /dev/null +++ b/data/review-receipts/papers/pal-code-reasoning-2022.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "pal-code-reasoning-2022", + "digest_sha256": "857cec22591d8950302786bf99210f37e3ce7a6b0fa364c8d4d2d4161dbae2fd" + }, + "source_revision": "arXiv:2211.10435v2", + "research_input_sha256": "3ca6a6ca60f5c6f71eb6b9d4938a302f25bb41767256c00cfa537682a58999ab", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/palm-2022.json b/data/review-receipts/papers/palm-2022.json new file mode 100644 index 000000000..704a5a99c --- /dev/null +++ b/data/review-receipts/papers/palm-2022.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "palm-2022", + "digest_sha256": "4a972542ca2f8b22240eeea1dbdb0c7228e32209323f41745c614a60d7ef76da" + }, + "source_revision": "arXiv:2204.02311v5", + "research_input_sha256": "f805c977a505eb3be69f5cf62d093a5b8612c7acb100dcf5c7c77f02c4262f71", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/plan-and-solve-prompting-2023.json b/data/review-receipts/papers/plan-and-solve-prompting-2023.json new file mode 100644 index 000000000..3b5091478 --- /dev/null +++ b/data/review-receipts/papers/plan-and-solve-prompting-2023.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "plan-and-solve-prompting-2023", + "digest_sha256": "873a84eb76b832aab87074cce0e2cfc58e56bb070e96ac8aa68547156cfceb85" + }, + "source_revision": "arXiv:2305.04091v3", + "research_input_sha256": "d2d4c19f25152b8a3d715afd4cce80015db445b28b4478b4f7e3be0c268393d9", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/prefix-tuning-2021.json b/data/review-receipts/papers/prefix-tuning-2021.json new file mode 100644 index 000000000..9993c4f9d --- /dev/null +++ b/data/review-receipts/papers/prefix-tuning-2021.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "prefix-tuning-2021", + "digest_sha256": "7a79e09d4d75a26bbd5ba80ee281206d64f9c05d13049a5097022a2fac458fe8" + }, + "source_revision": "arXiv:2101.00190v1", + "research_input_sha256": "5e1f594b5843377dc37b38b9203e2d1b735e1b43029c2176aff65db9047501ff", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/program-of-thoughts-2022.json b/data/review-receipts/papers/program-of-thoughts-2022.json new file mode 100644 index 000000000..b63109286 --- /dev/null +++ b/data/review-receipts/papers/program-of-thoughts-2022.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "program-of-thoughts-2022", + "digest_sha256": "3a1a76391b1207890db702882aa0d616d99582f37847bfd9ef646103a59e2023" + }, + "source_revision": "arXiv:2211.12588v4", + "research_input_sha256": "47ce46b950048baada6a41cc2ae0d641b81b7ff36e197babeb265893941ea0f2", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/prompt-to-prompt-2022.json b/data/review-receipts/papers/prompt-to-prompt-2022.json new file mode 100644 index 000000000..855911463 --- /dev/null +++ b/data/review-receipts/papers/prompt-to-prompt-2022.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "prompt-to-prompt-2022", + "digest_sha256": "90315253f92ad99c47c2a6fa10333e472969eba5ec09481ab3f6b12361d1927b" + }, + "source_revision": "arXiv:2208.01626v1", + "research_input_sha256": "fb544725021fffe4b3be6284890ee73dc150b5113c8b391cc1ff0660d1d50f99", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/prompt-tuning-2021.json b/data/review-receipts/papers/prompt-tuning-2021.json new file mode 100644 index 000000000..011b12948 --- /dev/null +++ b/data/review-receipts/papers/prompt-tuning-2021.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "prompt-tuning-2021", + "digest_sha256": "40305a0a2c27182badba740098a7116d259e98be14bcb9c8647496e99fc0f090" + }, + "source_revision": "arXiv:2104.08691v2", + "research_input_sha256": "d4f45f4b93bce04cff857847dc315f698383353fb23c00b95f3ea247cfb67bf1", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/qlora-2023.json b/data/review-receipts/papers/qlora-2023.json new file mode 100644 index 000000000..dfc614e66 --- /dev/null +++ b/data/review-receipts/papers/qlora-2023.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "qlora-2023", + "digest_sha256": "0e2421e11efb7222e959bc39379427ece730742793e14ced9c6848b41e79a6e5" + }, + "source_revision": "arXiv:2305.14314v1", + "research_input_sha256": "02545f5e4b90d560eda2a434b1a1e2f694bce57662d17a64440fe98f35f929d8", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/saycan-2022.json b/data/review-receipts/papers/saycan-2022.json new file mode 100644 index 000000000..62058ca7c --- /dev/null +++ b/data/review-receipts/papers/saycan-2022.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "saycan-2022", + "digest_sha256": "f9f36b3e3771766975e1dd4871bcf99a8b188701bbcdf7cecc104dd1c554e646" + }, + "source_revision": "arXiv:2204.01691v2", + "research_input_sha256": "6a4fb549e218f3e257706bea543d383d755b36b0bd6695cb350c91569cb34eaa", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/self-instruct-2022.json b/data/review-receipts/papers/self-instruct-2022.json new file mode 100644 index 000000000..a0776639b --- /dev/null +++ b/data/review-receipts/papers/self-instruct-2022.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "self-instruct-2022", + "digest_sha256": "968fb2595e4c994a2bbf9b3c56385b1b9018340ca33cb0a56e9f2114ac9fb6fe" + }, + "source_revision": "arXiv:2212.10560v2", + "research_input_sha256": "083aa5e217b02607a74d75cbbf9871d6c6f30c3b34cef4620ac938dc665be469", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/speculative-decoding-2022.json b/data/review-receipts/papers/speculative-decoding-2022.json new file mode 100644 index 000000000..7897de27f --- /dev/null +++ b/data/review-receipts/papers/speculative-decoding-2022.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "speculative-decoding-2022", + "digest_sha256": "1ba16b1b21153d74fd4acccb2e9acc65e3b82ad95d0584fdd5b296e22da85741" + }, + "source_revision": "arXiv:2211.17192v2", + "research_input_sha256": "d7e61964ae99d9b8912de0e1e166bda00dfc05c013daf4715f24c82bdcfb7c25", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/star-self-taught-reasoner-2022.json b/data/review-receipts/papers/star-self-taught-reasoner-2022.json new file mode 100644 index 000000000..1829cc7b9 --- /dev/null +++ b/data/review-receipts/papers/star-self-taught-reasoner-2022.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "star-self-taught-reasoner-2022", + "digest_sha256": "6905615a871b43a3b1d2872b16033f56ca41e5b8ff13d5b43380975173f73b78" + }, + "source_revision": "arXiv:2203.14465v2", + "research_input_sha256": "b96386383efe76f87d11da6ec72b86f7bf8aac6b0f699d721ff68f3f0ba97479", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/textual-inversion-2022.json b/data/review-receipts/papers/textual-inversion-2022.json new file mode 100644 index 000000000..29e93dc15 --- /dev/null +++ b/data/review-receipts/papers/textual-inversion-2022.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "textual-inversion-2022", + "digest_sha256": "7fe4f396ad1124ecae5664b8d7942a0a6d2acf82f80fd104b5733b70622359df" + }, + "source_revision": "arXiv:2208.01618v1", + "research_input_sha256": "15e6724d58d9ff5a0ab3591ea4727a3d79508a0a8dd7a49f906ad0fa8100e4af", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/toolllm-2023.json b/data/review-receipts/papers/toolllm-2023.json new file mode 100644 index 000000000..9b028c386 --- /dev/null +++ b/data/review-receipts/papers/toolllm-2023.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "toolllm-2023", + "digest_sha256": "a9d7ead607a2d552ece1b642c9b261aa14583395f923c67e2c57f4d244385215" + }, + "source_revision": "arXiv:2307.16789v2", + "research_input_sha256": "8c75cb0174d4816a8a8c321f8d0c80b769fa564438b5f2efca1bd4eaa7f4937b", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/toxigen-2022.json b/data/review-receipts/papers/toxigen-2022.json new file mode 100644 index 000000000..e9f05d56c --- /dev/null +++ b/data/review-receipts/papers/toxigen-2022.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "toxigen-2022", + "digest_sha256": "6864b9bb37134b3d24847401c2f5ee4d1a6491fd19703ba9c4a25ad94d100eb0" + }, + "source_revision": "arXiv:2203.09509v4", + "research_input_sha256": "6001f226a5327b067710c905af594b2e719efe1d9843d49ce4c2e09ca30e9be5", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/truthfulqa-2021.json b/data/review-receipts/papers/truthfulqa-2021.json new file mode 100644 index 000000000..ee04ef8b4 --- /dev/null +++ b/data/review-receipts/papers/truthfulqa-2021.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "truthfulqa-2021", + "digest_sha256": "1e9137fe2aaf875b54815039fe9d53aca9f7e1d332f25d2a00a25969f239f992" + }, + "source_revision": "arXiv:2109.07958v2", + "research_input_sha256": "31a0a9c012ace4fb2ab4cbb1ab6d210a79bd2281c1202dc9443f2a572ce11287", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/ul2-2022.json b/data/review-receipts/papers/ul2-2022.json new file mode 100644 index 000000000..24b1d6ce9 --- /dev/null +++ b/data/review-receipts/papers/ul2-2022.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "ul2-2022", + "digest_sha256": "bfdc46c05d78a872106fa71977a43b220f5e04681d9ecd60bff6e327025c6b4f" + }, + "source_revision": "arXiv:2205.05131v3", + "research_input_sha256": "f1fcd8b4c348fef6a4739e1ff75f9d99f036512b9173021876bc93cdeef873f3", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/webgpt-2021.json b/data/review-receipts/papers/webgpt-2021.json new file mode 100644 index 000000000..d1f558c0f --- /dev/null +++ b/data/review-receipts/papers/webgpt-2021.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "webgpt-2021", + "digest_sha256": "17eb3c519cde92d0915004103d77dbf5cda7a6cad3e9b8c6b05b85804de4dd78" + }, + "source_revision": "arXiv:2112.09332v3", + "research_input_sha256": "0d50a33d16b8002e8ee87f736887de6299bad7cb3ab984b0d46c046becd62ed9", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/data/review-receipts/papers/wizardlm-2023.json b/data/review-receipts/papers/wizardlm-2023.json new file mode 100644 index 000000000..4d22e12fa --- /dev/null +++ b/data/review-receipts/papers/wizardlm-2023.json @@ -0,0 +1,55 @@ +{ + "schema_version": "study-review-receipt-v1", + "generation": 1, + "predecessor_digest_sha256": null, + "note": { + "area": "papers", + "slug": "wizardlm-2023", + "digest_sha256": "cd4746f0aaeabc9658be82632cbefdc7765774d6c1a2df8f82a97aaa57f859c7" + }, + "source_revision": "arXiv:2304.12244v3", + "research_input_sha256": "c34bd2b904d86510d666641430bc3c9ef615dd9fe1a60da2af8fc927cfcbeb3d", + "reviewers": [ + { + "role": "ZERO_BASE", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 84, + "warnings": [ + "Zero-base explanation includes an analogy and toy mechanism check; paper benchmark was not executed." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "MANUAL_SIMULATION" + } + }, + { + "role": "ENGINEER", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 83, + "warnings": [ + "Claims are limited to arXiv metadata and static paper reading; no repository or service integration was run." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + }, + { + "role": "ACADEMIC", + "reviewer_version": "study-static-review-20260714-40-paper-batch", + "decision": "PASS_WITH_NOTES", + "score": 82, + "warnings": [ + "Citation identity was checked through arXiv metadata; reported experimental numbers are not independently reproduced." + ], + "execution": { + "review_mode": "STATIC_REVIEW", + "code_mode": "NOT_APPLICABLE" + } + } + ], + "waivers": [], + "created_at": "2026-07-14T12:00:00.000Z" +} diff --git a/src/content/docs/about.md b/src/content/docs/about.md index 7b83fd0a2..d0c762085 100644 --- a/src/content/docs/about.md +++ b/src/content/docs/about.md @@ -1,6 +1,6 @@ --- title: 关于这个站点 -description: 1900+ 篇精读笔记，按"未来工程师该懂什么"的标准筛选——给想成为 AI 时代工程师的同路人 +description: 2000+ 篇精读笔记，按"未来工程师该懂什么"的标准筛选——给想成为 AI 时代工程师的同路人 sidebar: order: 0 label: 立场宣言 @@ -14,7 +14,7 @@ sidebar: 写到今天的硬数字： -- **1023 篇论文笔记** + **961 篇项目笔记**，合计 **1900+ 篇** +- **1063 篇论文笔记** + **961 篇项目笔记**，合计 **2000+ 篇** - 横跨 19 个主题：分布式系统 76 / 编程语言 76 / 数据库 47 / 操作系统 46 / 机器学习 44 / 区块链 44 / 后端 API 40 / 基础设施 38 / 网络协议 37 / 图形学 36 / 形式化方法 27 / 通信 27 / 信息检索 25 / Agent 24 / CLI 23 / NLP 11 / 编译器 11 / 等 - 近 30 天集中产出：基础设施（444 commits）、编译器与 PL（72）、自演化 Agent（10+ 新建）、分布式（47）、区块链（44） @@ -166,7 +166,7 @@ AI 把"写代码"这件事的成本拉到了地板。但"在多个看似都对 - **编辑**：Jason 读、提观点、要求重写、调整声音、补判断 - **基础设施**：Astro + Starlight + GitHub Pages，由 Claude Code 搭 -1900+ 篇的产能不是单作者能写出来的——它是"AI 协作把人放大十倍"的活样本。藏着这件事反而会让站点失去价值；明说出来，你才能看到这种新工作方式怎么运转。 +2000+ 篇的产能不是单作者能写出来的——它是"AI 协作把人放大十倍"的活样本。藏着这件事反而会让站点失去价值；明说出来，你才能看到这种新工作方式怎么运转。 判断力部分必须由我贡献：**为什么是这个项目而不是它的同类**——这部分 AI 替不了。 diff --git a/src/content/docs/atlas/papers/topic-papers-agents-and-llm-systems-01.md b/src/content/docs/atlas/papers/topic-papers-agents-and-llm-systems-01.md index be122b175..5ae2471f7 100644 --- a/src/content/docs/atlas/papers/topic-papers-agents-and-llm-systems-01.md +++ b/src/content/docs/atlas/papers/topic-papers-agents-and-llm-systems-01.md @@ -1,6 +1,6 @@ --- title: "智能体与 LLM 系统 · 论文 · 第 1 组" -description: "68 条 智能体与 LLM 系统 Atlas 分块" +description: "100 条 智能体与 LLM 系统 Atlas 分块" sidebar: hidden: true --- @@ -9,7 +9,7 @@ sidebar: [返回论文全景索引](/study/papers-atlas/) -本分块共 68 条，稳定上限为 100 条。 +本分块共 100 条，稳定上限为 100 条。 | 论文 | Slug | 难度 | 可信状态 | 简介 | |---|---|---|---|---| @@ -22,6 +22,8 @@ sidebar: | [AutoGen — 多智能体对话框架](/study/papers/autogen/) | `autogen` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [AWQ — 看激活脸色给权重打折](/study/papers/awq/) | `awq` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [AWQ 2023 — 把 70B 大模型权重压到 35GB](/study/papers/awq-2023/) | `awq-2023` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | +| [BIG-Bench Hard — 从大题库里挑出模型最头疼的 23 类题](/study/papers/big-bench-hard-2022/) | `big-bench-hard-2022` | intermediate | UNVERIFIED | 用 BBH 理解为什么 benchmark 需要难题子集和 CoT 对照 | +| [BLOOM — 把 176B 多语种模型做成开放科学工程](/study/papers/bloom-2022/) | `bloom-2022` | intermediate | UNVERIFIED | 用 BLOOM 理解大模型也可以用社区协作、数据治理和开放发布来推进 | | [Chain-of-Thought — 让大模型先写步骤再回答](/study/papers/chain-of-thought/) | `chain-of-thought` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [ClawTrace — 把 agent 每步操作的"成本账"先算清再蒸馏](/study/papers/clawtrace-cost-aware/) | `clawtrace-cost-aware` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [Code as Agent Harness — 把代码当 agent 的"骨架"来重新看 agentic AI](/study/papers/code-as-agent-harness/) | `code-as-agent-harness` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | @@ -31,10 +33,17 @@ sidebar: | [EVE-Agent — 自我训练前先把证据钉在桌上](/study/papers/eve-agent-evidence/) | `eve-agent-evidence` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [Evo-Memory — 给"会自己长记性"的 agent 出一份统一考卷](/study/papers/evo-memory-2511/) | `evo-memory-2511` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [EXG 经验图 — 把 agent 的成败拼成一张可复用的关系图](/study/papers/exg-experience-graphs/) | `exg-experience-graphs` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | +| [Gorilla — 让 LLM 学会查 API 文档再调用](/study/papers/gorilla-2023/) | `gorilla-2023` | intermediate | UNVERIFIED | 用 Gorilla 理解 API grounding 如何降低工具调用幻觉 | | [GPTQ — 把 175B 大模型压成 4-bit 还几乎不掉点](/study/papers/gptq-2023/) | `gptq-2023` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [GraphRAG — 微软的知识图谱 + RAG](/study/papers/graphrag/) | `graphrag` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | +| [GSM8K — 小学数学题把大模型算术短板照出来](/study/papers/gsm8k-2021/) | `gsm8k-2021` | intermediate | UNVERIFIED | 用 GSM8K 理解数学 word problem、verifier 和采样重排为什么重要 | +| [HuggingGPT — 让 ChatGPT 当任务调度员，模型库当工具箱](/study/papers/hugginggpt-2023/) | `hugginggpt-2023` | intermediate | UNVERIFIED | 用 HuggingGPT 理解 LLM 如何规划并调用专用模型完成多模态任务 | +| [Inner Monologue — 让机器人把观察结果说回计划里](/study/papers/inner-monologue-2022/) | `inner-monologue-2022` | intermediate | UNVERIFIED | 用 Inner Monologue 理解闭环反馈如何让语言计划接上真实环境变化 | | [InstructGPT — RLHF 让 LLM 听话](/study/papers/instructgpt/) | `instructgpt` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | +| [Scaling Laws — 大模型训练不是玄学，是幂律预算题](/study/papers/kaplan-scaling-laws-2020/) | `kaplan-scaling-laws-2020` | advanced | UNVERIFIED | 用 Kaplan scaling laws 理解参数、数据和计算量怎样一起决定语言模型损失 | | [KV-Fold — 把 KV cache 当成 fold 的累加器，一段一段读长文](/study/papers/kv-fold/) | `kv-fold` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | +| [LaMDA — 聊天模型先学会有用、具体和不乱说](/study/papers/lamda-2022/) | `lamda-2022` | intermediate | UNVERIFIED | 用 LaMDA 理解开放域对话模型为什么需要质量、安全和 groundedness 三条线 | +| [Least-to-Most — 先拆小题，再解大题](/study/papers/least-to-most-prompting-2022/) | `least-to-most-prompting-2022` | intermediate | UNVERIFIED | 用 Least-to-Most Prompting 理解复杂推理为什么要先分解再逐步求解 | | [LLM.int8() — 大模型激活值里藏着几个超大异常通道](/study/papers/llm-int8-2022/) | `llm-int8-2022` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [LLM-Wiki — 把外部知识编译成 agent 自己的"维基"](/study/papers/llm-wiki-retrieval-reasoning/) | `llm-wiki-retrieval-reasoning` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [MCP-Bench — 用真实 MCP Server 测 agent 工具编排](/study/papers/mcp-bench-2025/) | `mcp-bench-2025` | intermediate | UNVERIFIED | MCP-Bench 通过 28 个 live MCP server 和 250 个工具评估多步工具编排 | @@ -44,25 +53,41 @@ sidebar: | [MemGym — 给长程 agent memory 做一间健身房](/study/papers/memgym/) | `memgym` | intermediate | UNVERIFIED | 用 MemGym 区分聊天记忆、执行记忆和可迁移的 agent 经验 | | [MetaGPT — 多智能体软件公司](/study/papers/metagpt/) | `metagpt` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [MIND-Skill — 用归纳和演绎双 agent 抽 skill 并保证质量](/study/papers/mind-skill/) | `mind-skill` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | +| [Minerva — 把语言模型拉进数学草稿纸](/study/papers/minerva-2022/) | `minerva-2022` | advanced | UNVERIFIED | 用 Minerva 理解为什么数学推理需要专门的数据、逐步解题和采样验证 | | [Misevolution — 自进化 agent 也会"越改越坏"，连顶配模型也躲不过](/study/papers/misevolution-2509/) | `misevolution-2509` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | +| [Mistral 7B — 小模型靠架构细节打出性价比](/study/papers/mistral-7b-2023/) | `mistral-7b-2023` | intermediate | UNVERIFIED | 用 Mistral 7B 理解 grouped-query attention 和 sliding-window attention 如何服务高效开源模型 | | [MMSkills — 把视觉 agent 的"操作经验"做成多模态卡片](/study/papers/mmskills-multimodal/) | `mmskills-multimodal` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | +| [MRKL — 给大模型配一组专家工具和路由器](/study/papers/mrkl-systems-2022/) | `mrkl-systems-2022` | intermediate | UNVERIFIED | 用 MRKL Systems 理解 neuro-symbolic agent 为什么要把 LLM、检索和计算模块拆开 | +| [Super-NaturalInstructions — 1600+ 任务教模型读懂说明书](/study/papers/natural-instructions-v2-2022/) | `natural-instructions-v2-2022` | intermediate | UNVERIFIED | 用 Super-NaturalInstructions 理解 declarative instructions 如何评测任务泛化 | | [NestedKV — 用三层记忆决定 KV cache 该留谁](/study/papers/nestedkv/) | `nestedkv` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [OpenHands — 开源 AI 软件工程师](/study/papers/openhands/) | `openhands` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | +| [OPT — 把 GPT-3 级训练日志打开给研究社区](/study/papers/opt-2022/) | `opt-2022` | intermediate | UNVERIFIED | 用 OPT 理解开放权重、训练日志和复现实验对 LLM 研究的重要性 | | [Orca — Transformer 生成模型的分布式推理调度](/study/papers/orca-2022/) | `orca-2022` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [Orca — 让一批 LLM 请求随到随走，不再排队等最长那个](/study/papers/orca-continuous-batching/) | `orca-continuous-batching` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | +| [Orca — 小模型不只抄答案，还学解释轨迹](/study/papers/orca-explanation-tuning-2023/) | `orca-explanation-tuning-2023` | intermediate | UNVERIFIED | 用 Orca 理解 explanation tuning 为什么比只蒸馏最终答案更像教学生 | | [OSCAR — 离线转个方向，把 KV Cache 压到 2-bit](/study/papers/oscar-int2-kv/) | `oscar-int2-kv` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [OSWorld — 把 GUI agent 放进真正的电脑里考试](/study/papers/osworld/) | `osworld` | intermediate | UNVERIFIED | 用 OSWorld 理解为什么电脑操作 agent 不能只在网页或脚本环境里评测 | +| [P-Tuning v2 — 把 prompt tuning 深插到每一层](/study/papers/p-tuning-v2-2021/) | `p-tuning-v2-2021` | intermediate | UNVERIFIED | 用 P-Tuning v2 理解深层连续提示为什么能跨规模和任务接近 fine-tuning | | [PagedAttention — 把 KV cache 当虚拟内存页来管理](/study/papers/paged-attention/) | `paged-attention` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [PagedAttention — 以页替代整段内存的显存管理](/study/papers/paged-attention-vllm/) | `paged-attention-vllm` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | +| [PAL — 让 Python 成为语言模型的草稿纸](/study/papers/pal-code-reasoning-2022/) | `pal-code-reasoning-2022` | intermediate | UNVERIFIED | 用 PAL 理解 Program-aided Language Models 如何把推理转成可运行代码 | +| [PaLM — Pathways 把 540B LLM 扩成统一底座](/study/papers/palm-2022/) | `palm-2022` | advanced | UNVERIFIED | 用 PaLM 理解大规模 dense decoder 如何在多任务、推理和代码能力上同时冒头 | +| [Plan-and-Solve — 零样本推理先写计划再执行](/study/papers/plan-and-solve-prompting-2023/) | `plan-and-solve-prompting-2023` | intermediate | UNVERIFIED | 用 Plan-and-Solve 理解为什么 prompt 可以显式拆成 plan 和 solve 两段 | +| [Prefix-Tuning — 不改模型，只给每层塞一段可训练前缀](/study/papers/prefix-tuning-2021/) | `prefix-tuning-2021` | intermediate | UNVERIFIED | 用 Prefix-Tuning 理解连续 prompt 如何成为参数高效微调方法 | +| [Program of Thoughts — 让模型写程序，把计算交给解释器](/study/papers/program-of-thoughts-2022/) | `program-of-thoughts-2022` | intermediate | UNVERIFIED | 用 Program of Thoughts 理解自然语言推理和精确计算为什么要分工 | +| [Prompt Tuning — 规模变大后，软提示也能接近微调](/study/papers/prompt-tuning-2021/) | `prompt-tuning-2021` | intermediate | UNVERIFIED | 用 Prompt Tuning 理解为什么 soft prompt 在大模型上突然变得有效 | +| [QLoRA — 4-bit 量化底座上贴 LoRA 也能微调](/study/papers/qlora-2023/) | `qlora-2023` | advanced | UNVERIFIED | 用 QLoRA 理解 NF4、double quantization 和 paged optimizers 如何降低微调门槛 | | [RAG (Lewis 2020) — 检索增强生成奠基](/study/papers/rag-lewis-2020/) | `rag-lewis-2020` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [ReAct — Reasoning and Acting](/study/papers/react/) | `react` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [ReAct Agent — 推理和行动交替的工具使用范式](/study/papers/react-agent/) | `react-agent` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [Reflexion — 让 LLM 自我反思](/study/papers/reflexion/) | `reflexion` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [RETRO — DeepMind 的检索增强 LLM](/study/papers/retro/) | `retro` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | +| [SayCan — 机器人不只问“想做什么”，还问“我能做什么”](/study/papers/saycan-2022/) | `saycan-2022` | intermediate | UNVERIFIED | 用 SayCan 理解语言模型和机器人 affordance 如何合成可执行动作 | | [Self-Consistency — 让模型把同一道题做 40 遍再投票](/study/papers/self-consistency-2022/) | `self-consistency-2022` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [自进化 AI agent 综述 — 给"会自己升级"的 agent 画一张统一地图](/study/papers/self-evolving-agents-survey/) | `self-evolving-agents-survey` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [Self-Evolving RecSys — 让 LLM agent 自己跑超参实验上线](/study/papers/self-evolving-recsys-2602/) | `self-evolving-recsys-2602` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [BDI-LLM Self-Evolving Agents — 让 agent 自己改自己源代码](/study/papers/self-evolving-software-agents/) | `self-evolving-software-agents` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | +| [Self-Instruct — 让模型自己造指令数据再学习](/study/papers/self-instruct-2022/) | `self-instruct-2022` | intermediate | UNVERIFIED | 用 Self-Instruct 理解指令微调数据如何从少量种子任务扩展出来 | | [SkCC — 给 LLM agent 写一个真正的 skill 编译器](/study/papers/skcc-skill-compiler/) | `skcc-skill-compiler` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [Skill-as-Pseudocode — 把 agent 笔记本写成可校验的伪代码](/study/papers/skill-as-pseudocode/) | `skill-as-pseudocode` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [Skill-Pro — 不动权重学可复用 skill 的非参数 PPO](/study/papers/skill-pro-nonparametric-ppo/) | `skill-pro-nonparametric-ppo` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | @@ -70,14 +95,23 @@ sidebar: | [SmoothQuant 2023 — 把激活的烫手山芋扔给权重](/study/papers/smoothquant-2023/) | `smoothquant-2023` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [SparseGPT — 175B 大模型一次过剪 50%，不重训](/study/papers/sparsegpt-2023/) | `sparsegpt-2023` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [SpecInfer — 让大模型一次"猜一棵树"再并行验证](/study/papers/specinfer-2023/) | `specinfer-2023` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | +| [Speculative Decoding — 小模型先猜，大模型只验收](/study/papers/speculative-decoding-2022/) | `speculative-decoding-2022` | advanced | UNVERIFIED | 用 Speculative Decoding 理解如何不改变分布地加速自回归生成 | +| [STaR — 模型先试着讲理由，再用对的理由训练自己](/study/papers/star-self-taught-reasoner-2022/) | `star-self-taught-reasoner-2022` | intermediate | UNVERIFIED | 用 STaR 理解 rationale bootstrapping 怎样减少人工推理标注 | | [SWE-Agent — Princeton SWE-bench 解法](/study/papers/swe-agent/) | `swe-agent` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [SWE-bench — 真实 GitHub Issue 评测](/study/papers/swe-bench/) | `swe-bench` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [SWE-Bench-CL — coding agent 不能只刷静态题](/study/papers/swe-bench-cl/) | `swe-bench-cl` | intermediate | UNVERIFIED | 用 SWE-Bench-CL 理解软件工程 agent 的持续学习、迁移和灾难性遗忘 | | [SWE-Skills-Bench — Agent 技能真的帮得上软件工程吗](/study/papers/swe-skills-bench-2026/) | `swe-skills-bench-2026` | intermediate | UNVERIFIED | 用 paired evaluation 衡量 SWE skills 对真实软件工程 agent 的边际收益和 token 成本 | | [ToolBench-X — 工具会坏时，agent 还能不能把事做完](/study/papers/toolbench-x/) | `toolbench-x` | intermediate | UNVERIFIED | 用 ToolBench-X 理解 tool-use benchmark 为什么要模拟规格漂移、调用错误、执行失败和结果冲突 | | [Toolformer — 教 LLM 自主调用 API](/study/papers/toolformer/) | `toolformer` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | +| [ToolLLM — 用 16000+ API 训练模型进入真实工具世界](/study/papers/toolllm-2023/) | `toolllm-2023` | intermediate | UNVERIFIED | 用 ToolLLM 理解大规模 API 数据集、工具检索和工具评测如何支撑 agent | +| [ToxiGen — 用生成模型造隐性仇恨测试集](/study/papers/toxigen-2022/) | `toxigen-2022` | intermediate | UNVERIFIED | 用 ToxiGen 理解安全评测为什么要覆盖隐性、对抗性和群体相关文本 | | [Tree of Thoughts — 让 LLM 像下棋一样多想几步再答](/study/papers/tree-of-thoughts-2023/) | `tree-of-thoughts-2023` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | +| [TruthfulQA — 专门问模型容易学人类谬误的问题](/study/papers/truthfulqa-2021/) | `truthfulqa-2021` | intermediate | UNVERIFIED | 用 TruthfulQA 理解语言模型为什么会模仿常见假话而不是坚持事实 | +| [UL2 — 一个模型同时练完补空、续写和长文本](/study/papers/ul2-2022/) | `ul2-2022` | advanced | UNVERIFIED | 用 UL2 理解 mixture-of-denoisers 如何统一不同语言模型训练范式 | | [VeriCache: Turning Lossy KV Cache into Lossless LLM Inference — 有损压缩草稿，无损输出验收](/study/papers/vericache/) | `vericache` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [Voyager — LLM 终身学习智能体](/study/papers/voyager/) | `voyager` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | +| [WebGPT — 让模型带着浏览器回答问题](/study/papers/webgpt-2021/) | `webgpt-2021` | intermediate | UNVERIFIED | 用 WebGPT 理解检索、引用和人类偏好如何组合成可追溯问答 | | [WebXSkill — 给 Web agent 的可执行 skill 是参数化代码 + URL 图索引](/study/papers/webxskill/) | `webxskill` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | -| [Zombie Agents — 自进化 agent 的长期记忆能被持久化"借尸还魂"](/study/papers/zombie-agents-2602/) | `zombie-agents-2602` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | +| [WizardLM — 用 Evol-Instruct 自动变难训练题](/study/papers/wizardlm-2023/) | `wizardlm-2023` | intermediate | UNVERIFIED | 用 WizardLM 理解 instruction 数据不只要多，还要逐步变复杂 | + +[下一组](/study/atlas/papers/topic-papers-agents-and-llm-systems-02/) diff --git a/src/content/docs/atlas/papers/topic-papers-agents-and-llm-systems-02.md b/src/content/docs/atlas/papers/topic-papers-agents-and-llm-systems-02.md new file mode 100644 index 000000000..1b9b9b8af --- /dev/null +++ b/src/content/docs/atlas/papers/topic-papers-agents-and-llm-systems-02.md @@ -0,0 +1,18 @@ +--- +title: "智能体与 LLM 系统 · 论文 · 第 2 组" +description: "1 条 智能体与 LLM 系统 Atlas 分块" +sidebar: + hidden: true +--- + + + +[返回论文全景索引](/study/papers-atlas/) + +本分块共 1 条，稳定上限为 100 条。 + +| 论文 | Slug | 难度 | 可信状态 | 简介 | +|---|---|---|---|---| +| [Zombie Agents — 自进化 agent 的长期记忆能被持久化"借尸还魂"](/study/papers/zombie-agents-2602/) | `zombie-agents-2602` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | + +[上一组](/study/atlas/papers/topic-papers-agents-and-llm-systems-01/) diff --git a/src/content/docs/atlas/papers/topic-papers-generative-models-and-diffusion-01.md b/src/content/docs/atlas/papers/topic-papers-generative-models-and-diffusion-01.md index b42002355..9a11b25bc 100644 --- a/src/content/docs/atlas/papers/topic-papers-generative-models-and-diffusion-01.md +++ b/src/content/docs/atlas/papers/topic-papers-generative-models-and-diffusion-01.md @@ -1,6 +1,6 @@ --- title: "生成模型 / 扩散 · 论文 · 第 1 组" -description: "10 条 生成模型 / 扩散 Atlas 分块" +description: "14 条 生成模型 / 扩散 Atlas 分块" sidebar: hidden: true --- @@ -9,17 +9,21 @@ sidebar: [返回论文全景索引](/study/papers-atlas/) -本分块共 10 条，稳定上限为 100 条。 +本分块共 14 条，稳定上限为 100 条。 | 论文 | Slug | 难度 | 可信状态 | 简介 | |---|---|---|---|---| +| [ControlNet — 给扩散模型加一条可控条件支路](/study/papers/controlnet-2023/) | `controlnet-2023` | intermediate | UNVERIFIED | 用 ControlNet 理解边缘、姿态和深度图如何稳定控制图像生成 | | [DALL-E 2 — 基于 CLIP + 扩散的图像生成](/study/papers/dalle-2/) | `dalle-2` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [DDIM — 把扩散模型 1000 步采样压到 50 步](/study/papers/ddim-2020/) | `ddim-2020` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [DDPM — Denoising Diffusion Probabilistic Models](/study/papers/ddpm/) | `ddpm` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [DiT — Diffusion Transformer](/study/papers/dit/) | `dit` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | +| [DreamBooth — 用几张图把一个新主体塞进生成模型](/study/papers/dreambooth-2022/) | `dreambooth-2022` | intermediate | UNVERIFIED | 用 DreamBooth 理解 subject-driven generation 怎样让扩散模型记住特定对象 | | [DreamFusion — 用 2D 扩散模型当老师，把 NeRF 教成 3D](/study/papers/dreamfusion-2022/) | `dreamfusion-2022` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [EDM — 把扩散模型的训练配方一次拆清楚](/study/papers/edm-2022/) | `edm-2022` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [LLaVA — 开源多模态对话模型](/study/papers/llava/) | `llava` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [Magic3D — 把 DreamFusion 的 NeRF 拆成"先粗后精"两阶段](/study/papers/magic3d-2023/) | `magic3d-2023` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [Parti — 把文生图当作翻译，用自回归 Transformer 一像素接一像素地写](/study/papers/parti-2022/) | `parti-2022` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | +| [Prompt-to-Prompt — 改词不改构图的 cross-attention 编辑](/study/papers/prompt-to-prompt-2022/) | `prompt-to-prompt-2022` | intermediate | UNVERIFIED | 用 Prompt-to-Prompt 理解扩散模型里文本 token 和图像布局如何对齐 | | [Stable Diffusion — 开源文生图引爆](/study/papers/stable-diffusion/) | `stable-diffusion` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | +| [Textual Inversion — 给新概念学一个专属 token](/study/papers/textual-inversion-2022/) | `textual-inversion-2022` | intermediate | UNVERIFIED | 用 Textual Inversion 理解冻结扩散模型时如何只学习概念 embedding | diff --git a/src/content/docs/atlas/papers/topic-papers-nlp-foundations-and-scaling-01.md b/src/content/docs/atlas/papers/topic-papers-nlp-foundations-and-scaling-01.md index 4d26f8d09..2151f1d88 100644 --- a/src/content/docs/atlas/papers/topic-papers-nlp-foundations-and-scaling-01.md +++ b/src/content/docs/atlas/papers/topic-papers-nlp-foundations-and-scaling-01.md @@ -21,6 +21,7 @@ sidebar: | [Attention Is All You Need](/study/papers/attention/) | `attention` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [Batch Normalization — 把每层激活值规整到 0 均值 1 方差，深网训练时间砍成 1/14](/study/papers/batchnorm-2015/) | `batchnorm-2015` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [BERT — 双向 Transformer 预训练](/study/papers/bert/) | `bert` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | +| [BigBird — 用稀疏 attention 拉长 Transformer 视野](/study/papers/bigbird-2020/) | `bigbird-2020` | advanced | UNVERIFIED | 用 BigBird 理解局部、全局和随机 attention 怎样组成长序列模式 | | [BigGAN — 把 GAN 暴力放大到 ImageNet 512×512](/study/papers/biggan-2018/) | `biggan-2018` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [BLIP-2 — 用 188M 小桥接器把冻结的视觉模型和大语言模型拼起来](/study/papers/blip2-2023/) | `blip2-2023` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [CCOPD — 让多轮对话别被自己的旧话带偏](/study/papers/ccopd-distillation/) | `ccopd-distillation` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | @@ -64,11 +65,13 @@ sidebar: | [Label Smoothing — 别让模型对正确答案过度自信](/study/papers/label-smoothing-2016/) | `label-smoothing-2016` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [Layer Normalization — 把归一化方向从 batch 转到 feature，让 RNN/Transformer 也能稳定训](/study/papers/layernorm-2016/) | `layernorm-2016` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [Linear Attention, Still: Why Mamba-style Models Plateau](/study/papers/linear-attention-still-2026/) | `linear-attention-still-2026` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | +| [Linformer — 把 attention 矩阵投影成线性复杂度](/study/papers/linformer-2020/) | `linformer-2020` | advanced | UNVERIFIED | 用 Linformer 理解低秩假设如何压缩 self-attention | | [Lion — 让程序自己搜出来的优化器，比 AdamW 内存少一半](/study/papers/lion-2023/) | `lion-2023` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [LLaMA — Meta 开源大语言模型](/study/papers/llama/) | `llama` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [LLMSurgeon — 从模型回答反推训练数据配方](/study/papers/llmsurgeon-data-mixture/) | `llmsurgeon-data-mixture` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [LoMo — 把同一句话换成图片也要看懂](/study/papers/lomo-modality/) | `lomo-modality` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [Longformer — 滑窗加少数全局 token，把长文档喂进 Transformer](/study/papers/longformer-2020/) | `longformer-2020` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | +| [LongNet — 用 dilated attention 把上下文推到十亿 token 想象空间](/study/papers/longnet-2023/) | `longnet-2023` | advanced | UNVERIFIED | 用 LongNet 理解扩张式 attention 如何在多尺度上连接超长序列 | | [Loong DocMT — 长文档翻译里的会挑上下文的代理](/study/papers/loong-doc-mt/) | `loong-doc-mt` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [LoRA — 给冻结大模型贴低秩便签](/study/papers/lora/) | `lora` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [彩票假设 — 大网里藏着一张能独立训出来的小网](/study/papers/lottery-ticket-2019/) | `lottery-ticket-2019` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | @@ -110,8 +113,5 @@ sidebar: | [SoundnessBench — 判断 AI 科学家会不会把坏点子当好点子](/study/papers/soundness-bench/) | `soundness-bench` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [StarCoder — 把训练数据完整公开的 15B 代码模型](/study/papers/starcoder-2023/) | `starcoder-2023` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [StyleGAN2 — 把 StyleGAN 的水滴瑕疵和潜空间纠葛一起修掉](/study/papers/stylegan2-2020/) | `stylegan2-2020` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | -| [Sycophancy 2023 — RLHF 模型为什么爱顺着用户说](/study/papers/sycophancy-2023/) | `sycophancy-2023` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | -| [T0 — 让 50 个人各写各的提示词，模型反而更会听新指令](/study/papers/t0-2021/) | `t0-2021` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | -| [T5 — Text-to-Text Transfer Transformer](/study/papers/t5/) | `t5` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | [下一组](/study/atlas/papers/topic-papers-nlp-foundations-and-scaling-02/) diff --git a/src/content/docs/atlas/papers/topic-papers-nlp-foundations-and-scaling-02.md b/src/content/docs/atlas/papers/topic-papers-nlp-foundations-and-scaling-02.md index 7ccac8dbc..f8b3f7e0d 100644 --- a/src/content/docs/atlas/papers/topic-papers-nlp-foundations-and-scaling-02.md +++ b/src/content/docs/atlas/papers/topic-papers-nlp-foundations-and-scaling-02.md @@ -1,6 +1,6 @@ --- title: "NLP 基础与 Scaling · 论文 · 第 2 组" -description: "11 条 NLP 基础与 Scaling Atlas 分块" +description: "14 条 NLP 基础与 Scaling Atlas 分块" sidebar: hidden: true --- @@ -9,10 +9,13 @@ sidebar: [返回论文全景索引](/study/papers-atlas/) -本分块共 11 条，稳定上限为 100 条。 +本分块共 14 条，稳定上限为 100 条。 | 论文 | Slug | 难度 | 可信状态 | 简介 | |---|---|---|---|---| +| [Sycophancy 2023 — RLHF 模型为什么爱顺着用户说](/study/papers/sycophancy-2023/) | `sycophancy-2023` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | +| [T0 — 让 50 个人各写各的提示词，模型反而更会听新指令](/study/papers/t0-2021/) | `t0-2021` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | +| [T5 — Text-to-Text Transfer Transformer](/study/papers/t5/) | `t5` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [TabPFN — 一秒解决小表格分类的 Transformer](/study/papers/tabpfn-2023/) | `tabpfn-2023` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [TD3 — 给 DDPG 装两副刹车，连续控制终于稳了](/study/papers/td3-2018/) | `td3-2018` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | | [Transformer — 让每个词一次看完整句话](/study/papers/transformer/) | `transformer` | unknown | UNVERIFIED | 暂无独立描述；可先从标题与正文定位开始。 | diff --git a/src/content/docs/career-plan.md b/src/content/docs/career-plan.md index acb2d2eae..ca5292e40 100644 --- a/src/content/docs/career-plan.md +++ b/src/content/docs/career-plan.md @@ -5,7 +5,7 @@ sidebar: order: 1 --- -> 本页是路径说明。具体笔记见左侧分组；当前规模 1900+ 篇（论文 1023 + 项目 961）。 +> 本页是路径说明。具体笔记见左侧分组；当前规模 2000+ 篇（论文 1063 + 项目 961）。 ## 1. 路径模型的演化 @@ -114,7 +114,7 @@ sidebar: ## 5. 当前优势与短板 -强项（截至当前 1900+ 篇规模）： +强项（截至当前 2000+ 篇规模）： - 编程语言与类型理论：完整覆盖 HM / λ / Hoare 链 - 分布式共识：Paxos / Raft / Lamport 主线齐全 diff --git a/src/content/docs/index.md b/src/content/docs/index.md index e201f8521..444efb49e 100644 --- a/src/content/docs/index.md +++ b/src/content/docs/index.md @@ -144,7 +144,7 @@ head: -

当前规模：1023 篇论文 + 961 个项目 = 1984 篇笔记，按 19 个主题组织。数量已移出首屏，只作为覆盖面证据。

+

当前规模：1063 篇论文 + 961 个项目 = 2024 篇笔记，按 19 个主题组织。数量已移出首屏，只作为覆盖面证据。

diff --git a/src/content/docs/method.md b/src/content/docs/method.md index 0ee10015b..f146d5427 100644 --- a/src/content/docs/method.md +++ b/src/content/docs/method.md @@ -1,6 +1,6 @@ --- title: 怎么消化一个 GitHub 项目 -description: 7 层方法论 + 1900 篇生产实践后的回看 +description: 7 层方法论 + 2000 篇生产实践后的回看 sidebar: order: 0 --- @@ -11,7 +11,7 @@ sidebar: ## 顶层结论（先看） -- 这套方法跑过 **961 篇项目 + 1023 篇论文 = 1984 篇**笔记，跨 19 个一级主题、约 1984 行写作密度 +- 这套方法跑过 **961 篇项目 + 1063 篇论文 = 2024 篇**笔记，跨 19 个一级主题、约 2024 行写作密度 - 最常被跳过的层是 **Layer 4 改一处**——但每次跳过都让整篇笔记从"机制"退回"翻译" - 真正变成"门面级"反向引用枢纽的笔记（[React](/study/projects/react/) 68 / [[pytorch]] 67 / [[kubernetes]] 66 / [[postgresql]] 66），无一例外都做过 L3+L4 双层 - L0 / L1 / L2 / L7 即使做得平庸也不致命；L3+L4 任一项缺失 = 整篇笔记掉档 @@ -238,7 +238,7 @@ sidebar: ## L4 改一处：从抽象到肌肉记忆的桥 -L4 是这套方法最被低估的一层。1900+ 篇后的具体观察： +L4 是这套方法最被低估的一层。2000+ 篇后的具体观察： - **改一处不是"做实验"，是"破除幻觉"**：READMEs 经常隐藏耦合点，改一行就暴露 - **3 类最高 ROI 的改动**： diff --git a/src/content/docs/papers-atlas.md b/src/content/docs/papers-atlas.md index 55cb70b7d..eb5279d9d 100644 --- a/src/content/docs/papers-atlas.md +++ b/src/content/docs/papers-atlas.md @@ -1,6 +1,6 @@ --- title: 论文全景索引 -description: 1023 篇论文的分块地图 · 稳定 taxonomy · 自动生成 +description: 1063 篇论文的分块地图 · 稳定 taxonomy · 自动生成 sidebar: order: 5 label: 论文全景索引 @@ -12,10 +12,10 @@ sidebar: ## 总览
-
1023论文总数
-
1004已有规范主题
+
1063论文总数
+
1044已有规范主题
19暂未收纳进主题路线
-
98.1%分类覆盖率（1004 / 1023，已分类 / 总数）
+
98.2%分类覆盖率（1044 / 1063，已分类 / 总数）
## 先选一条学习路径 @@ -41,10 +41,10 @@ Atlas 不替代精选路线。零基础读者先从下面六条路径选一条 | 主题 | English | 数量 | 分块 | |---|---|---:|---| -| 智能体与 LLM 系统 | Agents and LLM Systems | 68 | [第 1/1 组](/study/atlas/papers/topic-papers-agents-and-llm-systems-01/) | -| NLP 基础与 Scaling | NLP Foundations and Scaling | 111 | [第 1/2 组](/study/atlas/papers/topic-papers-nlp-foundations-and-scaling-01/) · [第 2/2 组](/study/atlas/papers/topic-papers-nlp-foundations-and-scaling-02/) | +| 智能体与 LLM 系统 | Agents and LLM Systems | 101 | [第 1/2 组](/study/atlas/papers/topic-papers-agents-and-llm-systems-01/) · [第 2/2 组](/study/atlas/papers/topic-papers-agents-and-llm-systems-02/) | +| NLP 基础与 Scaling | NLP Foundations and Scaling | 114 | [第 1/2 组](/study/atlas/papers/topic-papers-nlp-foundations-and-scaling-01/) · [第 2/2 组](/study/atlas/papers/topic-papers-nlp-foundations-and-scaling-02/) | | 计算机视觉 | Computer Vision | 11 | [第 1/1 组](/study/atlas/papers/topic-papers-computer-vision-01/) | -| 生成模型 / 扩散 | Generative Models and Diffusion | 10 | [第 1/1 组](/study/atlas/papers/topic-papers-generative-models-and-diffusion-01/) | +| 生成模型 / 扩散 | Generative Models and Diffusion | 14 | [第 1/1 组](/study/atlas/papers/topic-papers-generative-models-and-diffusion-01/) | | 强化学习 | Reinforcement Learning | 8 | [第 1/1 组](/study/atlas/papers/topic-papers-reinforcement-learning-01/) | | AI 安全与可解释性 | AI Safety and Interpretability | 9 | [第 1/1 组](/study/atlas/papers/topic-papers-ai-safety-and-interpretability-01/) | | 分布式系统 | Distributed Systems | 104 | [第 1/2 组](/study/atlas/papers/topic-papers-distributed-systems-01/) · [第 2/2 组](/study/atlas/papers/topic-papers-distributed-systems-02/) | @@ -79,4 +79,4 @@ Atlas 不替代精选路线。零基础读者先从下面六条路径选一条 - difficulty 未知：1014 - description 为空：1013 -- sidecar 主键：1023 个唯一 `area::slug` +- sidecar 主键：1063 个唯一 `area::slug` diff --git a/src/content/docs/papers-method.md b/src/content/docs/papers-method.md index de7b80dc5..d5665bf9a 100644 --- a/src/content/docs/papers-method.md +++ b/src/content/docs/papers-method.md @@ -11,7 +11,7 @@ sidebar: ## 站点的论文体量 -截至 2026-07，论文目录共 1023 篇笔记，覆盖： +截至 2026-07，论文目录共 1063 篇笔记，覆盖： - 分布式系统 76 篇（[[paxos-1998]] / [[raft]] / [[lamport-1978]] / [[spanner-2012]]） - 编程语言 + 类型论 76 篇（[[hindley-milner]] / [[lambda-calculus]] / [[hoare-logic]]） diff --git a/src/content/docs/papers-queue.md b/src/content/docs/papers-queue.md index e081e4bbf..c8420a69e 100644 --- a/src/content/docs/papers-queue.md +++ b/src/content/docs/papers-queue.md @@ -1,11 +1,11 @@ --- title: 论文队列 -description: 按 topic 分组的 pillar 推荐 —— 站内 1023 篇论文笔记里，每条主线挑 3-5 篇代表作做切入点 +description: 按 topic 分组的 pillar 推荐 —— 站内 1063 篇论文笔记里，每条主线挑 3-5 篇代表作做切入点 sidebar: order: 4 --- -> 站内累计 1023 篇论文笔记，跨 14 个主题。这页不是"待读清单"，是 +> 站内累计 1063 篇论文笔记，跨 14 个主题。这页不是"待读清单"，是 > **入门指引** —— 每个 topic 给 3-5 篇 pillar 论文 + 一行说明它 > 为什么是该 topic 的支点。看完一条主线的 pillar，你就拿到了 > 该 topic 整张反向链接图的入口。 @@ -13,7 +13,7 @@ sidebar: ## 怎么用这页 - 不知道某个 topic 从哪读 → 来这里挑该主题 3-5 篇 pillar -- 想看完整 1023 篇分布与主题地图 → [papers-atlas](/study/papers-atlas/) +- 想看完整 1063 篇分布与主题地图 → [papers-atlas](/study/papers-atlas/) - 想要"如何精读一篇论文"的方法 → [papers-method](/study/papers-method/) - 想看跨论文 + 项目的混合阅读节奏 → [queue](/study/queue/) diff --git a/src/content/docs/papers/big-bench-hard-2022.md b/src/content/docs/papers/big-bench-hard-2022.md new file mode 100644 index 000000000..4f82f7c9c --- /dev/null +++ b/src/content/docs/papers/big-bench-hard-2022.md @@ -0,0 +1,90 @@ +--- +title: 'BIG-Bench Hard — 从大题库里挑出模型最头疼的 23 类题' +description: '用 BBH 理解为什么 benchmark 需要难题子集和 CoT 对照。' +来源: 'Suzgun et al., arXiv:2210.09261' +日期: 2026-07-14 +分类: LLM / Evaluation +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2210.09261v1 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2210.09261 + source_version: arXiv:2210.09261v1 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v1 +--- + +## 是什么 + +Challenging BIG-Bench Tasks and Whether Chain-of-Thought Can Solve Them 是一篇 LLM / Evaluation 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像从整本习题集里挑出全班错误率最高的题，专门看模型是不是真会推理。 + +它在本轮 40 篇里的位置是 **Batch 10 / evaluation and safety**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +BIG-bench 很大，但平均分会掩盖模型最薄弱的任务。研究者需要一个更聚焦的难题集合。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| Hard subset | 挑出模型表现低于人类的 23 个任务。 | +| CoT 对照 | 比较普通 prompting 和 chain-of-thought。 | +| 多任务诊断 | 覆盖逻辑、符号、常识和多步推理。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +如果总题库 200 道平均 80 分，但 23 道逻辑题只有 30 分，BBH 就把这 23 道拿出来单独追踪。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **难题子集会被反复优化**：难题子集会被反复优化，长期需要更新。 +2. **CoT 提升不代表推理机制完全可靠。**：CoT 提升不代表推理机制完全可靠。 +3. **任务格式仍是文本题**：任务格式仍是文本题，不能代表工具和交互 agent。 +4. **难度选择依赖当时模型水平**：难度选择依赖当时模型水平，强模型时代要重估。 + +## 学到什么 + +- 评测要看弱点集合，而不是只看大平均分。 +- BBH 是 CoT 时代最常见的推理诊断集合之一。 +- 产品验收也应该保留“最容易失败的固定小集”。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[bigbench-2022]]、[[chain-of-thought]]、[[least-to-most-prompting-2022]]、[[agent-planning-benchmark-2026]] + +## 关联 + +- [[bigbench-2022]] +- [[chain-of-thought]] +- [[least-to-most-prompting-2022]] +- [[agent-planning-benchmark-2026]] + +## 反向链接 + + diff --git a/src/content/docs/papers/bigbird-2020.md b/src/content/docs/papers/bigbird-2020.md new file mode 100644 index 000000000..d6a95cbfa --- /dev/null +++ b/src/content/docs/papers/bigbird-2020.md @@ -0,0 +1,90 @@ +--- +title: 'BigBird — 用稀疏 attention 拉长 Transformer 视野' +description: '用 BigBird 理解局部、全局和随机 attention 怎样组成长序列模式。' +来源: 'Zaheer et al., arXiv:2007.14062' +日期: 2026-07-14 +分类: NLP / Efficient Attention +难度: 高级 +difficulty: advanced +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2007.14062v2 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2007.14062 + source_version: arXiv:2007.14062v2 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v2 +--- + +## 是什么 + +Big Bird: Transformers for Longer Sequences 是一篇 Transformer / Efficient Attention 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像开会时不需要每个人和每个人都私聊：相邻同事聊局部，主持人做全局，随机跨组交流补信息。 + +它在本轮 40 篇里的位置是 **Batch 8 / long context and inference**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +长文档、基因序列等任务需要上千到上万 token，标准 attention 太贵。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| 局部窗口 | 每个 token 看附近 token。 | +| 全局 token | 少数特殊 token 连接全局信息。 | +| 随机连接 | 补充远距离信息路径，保持表达能力。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +读一本长报告时，每段先看前后段，目录页提供全局索引，再随机抽查远处引用，成本远低于所有段落互相比较。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **稀疏模式是先验**：稀疏模式是先验，选错模式会漏掉关键依赖。 +2. **理论表达能力不等于具体任务效果。**：理论表达能力不等于具体任务效果。 +3. **全局 token 设计会影响信息汇聚。**：全局 token 设计会影响信息汇聚。 +4. **实现效率依赖 kernel 和硬件支持。**：实现效率依赖 kernel 和硬件支持。 + +## 学到什么 + +- BigBird 展示了“稀疏但连通”的 attention 设计哲学。 +- 长上下文不是只把窗口拉大，还要设计信息路由。 +- 后续 LongNet、Gemini 1.5 等都在不同层面延续这个问题。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[longformer-2020]]、[[linformer-2020]]、[[longnet-2023]]、[[gemini-1.5-2024]] + +## 关联 + +- [[longformer-2020]] +- [[linformer-2020]] +- [[longnet-2023]] +- [[gemini-1.5-2024]] + +## 反向链接 + + diff --git a/src/content/docs/papers/bloom-2022.md b/src/content/docs/papers/bloom-2022.md new file mode 100644 index 000000000..c14e6539d --- /dev/null +++ b/src/content/docs/papers/bloom-2022.md @@ -0,0 +1,90 @@ +--- +title: 'BLOOM — 把 176B 多语种模型做成开放科学工程' +description: '用 BLOOM 理解大模型也可以用社区协作、数据治理和开放发布来推进。' +来源: 'BigScience Workshop, arXiv:2211.05100' +日期: 2026-07-14 +分类: LLM / Open Science +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2211.05100v4 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2211.05100 + source_version: arXiv:2211.05100v4 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v4 +--- + +## 是什么 + +BLOOM: A 176B-Parameter Open-Access Multilingual Language Model 是一篇 LLM / Open Science 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像几百人一起修一座公共图书馆：书从哪里来、有哪些语言、谁能进馆，都要写清楚。 + +它在本轮 40 篇里的位置是 **Batch 2 / open and dialogue models**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +大模型训练往往由少数公司闭门完成，多语种覆盖和数据来源透明度不足，研究者难以审计偏差。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| BigScience 协作 | 用开放工作组组织模型、数据、法律和伦理工作。 | +| ROOTS 语料 | 为多语种训练整理来源和治理记录。 | +| 开放访问 | 发布模型和文档，让社区复查与再利用。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +如果一个模型号称会 40 种语言，但不公开各语言数据比例，你无法判断低资源语言表现差是模型问题还是数据问题。BLOOM 把这类问题前移到数据卡和治理流程。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **开放科学不自动消除偏见**：开放科学不自动消除偏见，只是让偏见更容易被看见。 +2. **多语种覆盖不是平均能力**：多语种覆盖不是平均能力，数据量和质量仍高度不均。 +3. **176B 的开放访问仍有硬件门槛**：176B 的开放访问仍有硬件门槛，推理不是人人可跑。 +4. **协作治理成本很高**：协作治理成本很高，不能只按模型分数评价项目。 + +## 学到什么 + +- BLOOM 的核心价值是透明过程和多语种公共资产。 +- 开放模型需要数据、许可证、模型卡和访问政策一起设计。 +- 它给后来的开源 LLM 生态提供了组织范式。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[opt-2022]]、[[llama]]、[[mistral-7b-2023]]、[[gpt-3]] + +## 关联 + +- [[opt-2022]] +- [[llama]] +- [[mistral-7b-2023]] +- [[gpt-3]] + +## 反向链接 + + diff --git a/src/content/docs/papers/controlnet-2023.md b/src/content/docs/papers/controlnet-2023.md new file mode 100644 index 000000000..8969a32a0 --- /dev/null +++ b/src/content/docs/papers/controlnet-2023.md @@ -0,0 +1,90 @@ +--- +title: 'ControlNet — 给扩散模型加一条可控条件支路' +description: '用 ControlNet 理解边缘、姿态和深度图如何稳定控制图像生成。' +来源: 'Zhang et al., arXiv:2302.05543' +日期: 2026-07-14 +分类: Diffusion / Control +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2302.05543v3 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2302.05543 + source_version: arXiv:2302.05543v3 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v3 +--- + +## 是什么 + +Adding Conditional Control to Text-to-Image Diffusion Models 是一篇 Diffusion / Control 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像画师已经会画风格，现在给他一张铅笔草图，要求构图必须跟草图走。 + +它在本轮 40 篇里的位置是 **Batch 9 / controllable generation**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +纯文本控制扩散模型太松，用户想固定姿态、边缘、深度或布局时，prompt 很难精确约束。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| 锁定原模型 | 保留预训练 diffusion backbone 的生成能力。 | +| 可训练条件分支 | 为边缘、深度、pose 等条件学习控制信号。 | +| Zero convolution | 让新分支从不破坏原模型开始逐渐学习。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +给一张 Canny 边缘图和 prompt“水彩风房子”，ControlNet 会沿着边缘图生成，而不是自由发挥构图。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **条件图质量决定结果上限**：条件图质量决定结果上限，坏边缘会带来坏生成。 +2. **控制强度过高会牺牲多样性。**：控制强度过高会牺牲多样性。 +3. **不同条件类型需要不同训练数据。**：不同条件类型需要不同训练数据。 +4. **版权和肖像问题不会因可控生成自动消失。**：版权和肖像问题不会因可控生成自动消失。 + +## 学到什么 + +- 可控生成的关键是把用户意图从文本扩展到结构化条件。 +- ControlNet 让扩散模型从玩具出图更接近设计工具。 +- 它说明“冻结强底座 + 训练控制支路”是高性价比路线。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[ddpm]]、[[edm-2022]]、[[dreambooth-2022]]、[[prompt-to-prompt-2022]] + +## 关联 + +- [[ddpm]] +- [[edm-2022]] +- [[dreambooth-2022]] +- [[prompt-to-prompt-2022]] + +## 反向链接 + + diff --git a/src/content/docs/papers/dreambooth-2022.md b/src/content/docs/papers/dreambooth-2022.md new file mode 100644 index 000000000..0cf24fe96 --- /dev/null +++ b/src/content/docs/papers/dreambooth-2022.md @@ -0,0 +1,90 @@ +--- +title: 'DreamBooth — 用几张图把一个新主体塞进生成模型' +description: '用 DreamBooth 理解 subject-driven generation 怎样让扩散模型记住特定对象。' +来源: 'Ruiz et al., arXiv:2208.12242' +日期: 2026-07-14 +分类: Diffusion / Personalization +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2208.12242v2 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2208.12242 + source_version: arXiv:2208.12242v2 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v2 +--- + +## 是什么 + +DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation 是一篇 Diffusion / Personalization 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像让画师看几张你家杯子的照片，然后能把同一个杯子画到海边、办公室和油画风场景里。 + +它在本轮 40 篇里的位置是 **Batch 9 / controllable generation**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +文本到图像模型知道“狗”或“背包”，但不知道用户指定的那一只狗或那一个背包。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| 稀有 token 绑定主体 | 用特殊词指代新主体。 | +| 少样本微调 | 用几张主体图片调整模型。 | +| Prior preservation | 防止模型把整个类别都过拟合成这个主体。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +输入 5 张同一只玩具熊照片，学习 token `sks bear`，之后 prompt“sks bear wearing sunglasses”生成同主体新场景。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **过拟合会让主体只能复刻训练照片姿势。**：过拟合会让主体只能复刻训练照片姿势。 +2. **prior preservation 不足会污染通用类别。**：prior preservation 不足会污染通用类别。 +3. **个人主体生成涉及肖像权和授权边界。**：个人主体生成涉及肖像权和授权边界。 +4. **微调成本比 Textual Inversion 更高。**：微调成本比 Textual Inversion 更高。 + +## 学到什么 + +- 个性化生成需要在“记住主体”和“保留模型常识”之间平衡。 +- DreamBooth 是生成式产品商业化的重要技术节点。 +- 少样本定制越强，滥用和版权治理越重要。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[textual-inversion-2022]]、[[controlnet-2023]]、[[ddpm]]、[[edm-2022]] + +## 关联 + +- [[textual-inversion-2022]] +- [[controlnet-2023]] +- [[ddpm]] +- [[edm-2022]] + +## 反向链接 + + diff --git a/src/content/docs/papers/gorilla-2023.md b/src/content/docs/papers/gorilla-2023.md new file mode 100644 index 000000000..cd7028b61 --- /dev/null +++ b/src/content/docs/papers/gorilla-2023.md @@ -0,0 +1,90 @@ +--- +title: 'Gorilla — 让 LLM 学会查 API 文档再调用' +description: '用 Gorilla 理解 API grounding 如何降低工具调用幻觉。' +来源: 'Patil et al., arXiv:2305.15334' +日期: 2026-07-14 +分类: LLM / Tool Use +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2305.15334v1 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2305.15334 + source_version: arXiv:2305.15334v1 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v1 +--- + +## 是什么 + +Gorilla: Large Language Model Connected with Massive APIs 是一篇 LLM / Tool Use 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像程序员写代码前先查官方文档，而不是凭记忆猜函数名和参数。 + +它在本轮 40 篇里的位置是 **Batch 6 / agent tool ecosystems**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +LLM 调 API 时常编造不存在的函数、参数或版本。工具越多，幻觉空间越大。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| APIBench | 整理大量机器学习 API 调用任务。 | +| 检索增强 | 先找相关 API 文档，再生成调用。 | +| 调用格式评测 | 检查函数名、参数和版本是否真实可用。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +用户要加载 Hugging Face 模型。Gorilla 式流程先检索 `transformers.pipeline` 文档，再输出参数，而不是凭模型记忆写一个不存在的 `load_hf_model()`。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **检索到旧版本文档会导致过时调用。**：检索到旧版本文档会导致过时调用。 +2. **API 调用正确不等于业务流程正确。**：API 调用正确不等于业务流程正确。 +3. **长尾库文档质量差时**：长尾库文档质量差时，模型仍可能猜。 +4. **真实生产还要处理鉴权、速率限制和错误返回。**：真实生产还要处理鉴权、速率限制和错误返回。 + +## 学到什么 + +- 工具调用可靠性首先是文档 grounding 问题。 +- Gorilla 把“会用工具”从自然语言能力转成 API 版本契约。 +- MCP、function calling 和 tool benchmark 都需要类似检查。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[toolllm-2023]]、[[toolformer]]、[[mcpworld-2025]]、[[mrkl-systems-2022]] + +## 关联 + +- [[toolllm-2023]] +- [[toolformer]] +- [[mcpworld-2025]] +- [[mrkl-systems-2022]] + +## 反向链接 + + diff --git a/src/content/docs/papers/gsm8k-2021.md b/src/content/docs/papers/gsm8k-2021.md new file mode 100644 index 000000000..a951a25f3 --- /dev/null +++ b/src/content/docs/papers/gsm8k-2021.md @@ -0,0 +1,90 @@ +--- +title: 'GSM8K — 小学数学题把大模型算术短板照出来' +description: '用 GSM8K 理解数学 word problem、verifier 和采样重排为什么重要。' +来源: 'Cobbe et al., arXiv:2110.14168' +日期: 2026-07-14 +分类: LLM / Math Evaluation +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2110.14168v2 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2110.14168 + source_version: arXiv:2110.14168v2 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v2 +--- + +## 是什么 + +Training Verifiers to Solve Math Word Problems 是一篇 LLM / Math Evaluation 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像小学应用题：文字都看得懂，但一步漏算、单位错了，答案就错。 + +它在本轮 40 篇里的位置是 **Batch 10 / evaluation and safety**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +LLM 在自然语言上强，但多步算术 word problem 容易出现中间步骤错误。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| GSM8K 数据集 | 整理高质量小学数学文字题和解答。 | +| 多样本生成 | 让模型生成多个候选解。 | +| Verifier 重排 | 训练验证器从候选答案中挑更可信的一条。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +同一题生成 20 个解法，有些列式对、有些算错。verifier 学会偏好步骤自洽且答案正确的候选。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **verifier 也可能学到表面模式**：verifier 也可能学到表面模式，不能保证证明正确。 +2. **小学数学不覆盖高等数学、代码和真实数据分析。**：小学数学不覆盖高等数学、代码和真实数据分析。 +3. **多样本采样提高成本。**：多样本采样提高成本。 +4. **训练集污染会显著影响数学 benchmark。**：训练集污染会显著影响数学 benchmark。 + +## 学到什么 + +- GSM8K 把数学推理评测变成 LLM 标配。 +- “生成多个候选 + 验证器选择”是推理可靠性的通用结构。 +- 后续 Minerva、PAL、Program of Thoughts 都在围绕它扩展。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[minerva-2022]]、[[program-of-thoughts-2022]]、[[pal-code-reasoning-2022]]、[[self-consistency-2022]] + +## 关联 + +- [[minerva-2022]] +- [[program-of-thoughts-2022]] +- [[pal-code-reasoning-2022]] +- [[self-consistency-2022]] + +## 反向链接 + + diff --git a/src/content/docs/papers/hugginggpt-2023.md b/src/content/docs/papers/hugginggpt-2023.md new file mode 100644 index 000000000..a4e16df18 --- /dev/null +++ b/src/content/docs/papers/hugginggpt-2023.md @@ -0,0 +1,90 @@ +--- +title: 'HuggingGPT — 让 ChatGPT 当任务调度员，模型库当工具箱' +description: '用 HuggingGPT 理解 LLM 如何规划并调用专用模型完成多模态任务。' +来源: 'Shen et al., arXiv:2303.17580' +日期: 2026-07-14 +分类: LLM / Tool Orchestration +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2303.17580v4 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2303.17580 + source_version: arXiv:2303.17580v4 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v4 +--- + +## 是什么 + +HuggingGPT: Solving AI Tasks with ChatGPT and its Friends in Hugging Face 是一篇 LLM / Tool Orchestration 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像项目经理接到需求后，把抠图、翻译、语音识别分别派给专业同事，最后合并交付。 + +它在本轮 40 篇里的位置是 **Batch 6 / agent tool ecosystems**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +单个 LLM 不擅长所有模态和专业任务，但模型社区已经有大量专用模型，缺少统一调度层。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| 任务规划 | LLM 把用户请求拆成多个子任务。 | +| 模型选择 | 从 Hugging Face 模型描述中选合适工具。 | +| 执行与汇总 | 调用模型、收集结果，再生成最终回答。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +用户上传图片并要求“描述图片，再生成一段配乐提示”。系统先调用图像描述模型，再把文本交给音乐/文本生成模型。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **模型描述不等于能力保证**：模型描述不等于能力保证，选择错工具会级联失败。 +2. **多模型流水线的延迟和费用会快速累加。**：多模型流水线的延迟和费用会快速累加。 +3. **中间结果格式不统一**：中间结果格式不统一，编排层要做适配。 +4. **开源模型许可证和安全风险不能被调度层忽略。**：开源模型许可证和安全风险不能被调度层忽略。 + +## 学到什么 + +- HuggingGPT 把 LLM 定位成 orchestrator，而不是万能执行器。 +- 工具生态越大，模型选择、状态传递和错误恢复越重要。 +- 今天的 agent workflow 平台仍在解决这套编排问题。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[mrkl-systems-2022]]、[[gorilla-2023]]、[[toolllm-2023]]、[[mcp-bench-2025]] + +## 关联 + +- [[mrkl-systems-2022]] +- [[gorilla-2023]] +- [[toolllm-2023]] +- [[mcp-bench-2025]] + +## 反向链接 + + diff --git a/src/content/docs/papers/inner-monologue-2022.md b/src/content/docs/papers/inner-monologue-2022.md new file mode 100644 index 000000000..5d7c024a8 --- /dev/null +++ b/src/content/docs/papers/inner-monologue-2022.md @@ -0,0 +1,90 @@ +--- +title: 'Inner Monologue — 让机器人把观察结果说回计划里' +description: '用 Inner Monologue 理解闭环反馈如何让语言计划接上真实环境变化。' +来源: 'Huang et al., arXiv:2207.05608' +日期: 2026-07-14 +分类: LLM Agent / Robotics +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2207.05608v1 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2207.05608 + source_version: arXiv:2207.05608v1 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v1 +--- + +## 是什么 + +Inner Monologue: Embodied Reasoning through Planning with Language Models 是一篇 Embodied AI / Robotics 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像搬家时一边做一边自言自语：“箱子太重，先找推车；门关着，先开门。”这些反馈会改变下一步。 + +它在本轮 40 篇里的位置是 **Batch 5 / agents and tools**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +只生成一次性计划的机器人容易在环境变化、动作失败或目标不清时卡住。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| 语言化反馈 | 把视觉、成功/失败和环境状态转成文本。 | +| 循环计划 | LLM 根据新反馈继续生成下一步。 | +| 多来源上下文 | 把人类指令、机器人状态和观察合并进 prompt。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +机器人计划“拿杯子”，执行后反馈“抓取失败，杯子太远”。下一轮计划变成“移动到桌边再抓取”，而不是重复失败动作。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **反馈文本如果不准**：反馈文本如果不准，会把模型带偏。 +2. **循环越长**：循环越长，prompt 越容易积累噪声。 +3. **语言化观察会丢失细节**：语言化观察会丢失细节，不能替代底层控制和感知。 +4. **失败恢复需要策略边界**：失败恢复需要策略边界，不能无限重试。 + +## 学到什么 + +- agent 的“内心独白”本质是状态回流机制。 +- 闭环比一次性计划更接近真实机器人和电脑操作。 +- 软件工作流里的日志、测试、截图也可以看作 inner monologue。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[saycan-2022]]、[[react-agent]]、[[osworld]]、[[agent-planning-benchmark-2026]] + +## 关联 + +- [[saycan-2022]] +- [[react-agent]] +- [[osworld]] +- [[agent-planning-benchmark-2026]] + +## 反向链接 + + diff --git a/src/content/docs/papers/kaplan-scaling-laws-2020.md b/src/content/docs/papers/kaplan-scaling-laws-2020.md new file mode 100644 index 000000000..d87b20083 --- /dev/null +++ b/src/content/docs/papers/kaplan-scaling-laws-2020.md @@ -0,0 +1,90 @@ +--- +title: 'Scaling Laws — 大模型训练不是玄学，是幂律预算题' +description: '用 Kaplan scaling laws 理解参数、数据和计算量怎样一起决定语言模型损失。' +来源: 'Kaplan et al., arXiv:2001.08361' +日期: 2026-07-14 +分类: LLM / Scaling Laws +难度: 高级 +difficulty: advanced +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2001.08361v1 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2001.08361 + source_version: arXiv:2001.08361v1 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v1 +--- + +## 是什么 + +Scaling Laws for Neural Language Models 是一篇 LLM / Scaling Laws 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像给一家工厂做产能规划：机器太多但原料不够会空转，原料太多但机器太少也堆仓库，训练 LLM 也要在参数、数据和算力之间找平衡。 + +它在本轮 40 篇里的位置是 **Batch 1 / foundation scaling**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +论文出现前，大家知道“大模型通常更好”，但不知道多大、多长数据、多少计算量之间该怎么配。没有这张预算图，训练计划很容易靠经验拍脑袋。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| 幂律拟合 | 把 loss 和模型规模、数据量、计算量之间的关系写成稳定曲线。 | +| 计算最优边界 | 在固定 compute 下找参数量和 token 数的配比。 | +| 跨尺度外推 | 用小模型实验估算大模型训练会落在哪个区间。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +如果预算只能训练 10 天，一个 10B 模型只看 1B token 可能欠训练；一个 100M 模型看 1T token 又容量不够。scaling law 的 toy 复现就是画出两条 loss 曲线，找“再加参数”和“再加数据”边际收益相近的位置。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **幂律不是物理定律**：幂律不是物理定律，换数据分布、架构或优化器后要重新校准。 +2. **论文早期结论偏向“参数多、数据相对少”**：论文早期结论偏向“参数多、数据相对少”，后来 Chinchilla 修正了 compute-optimal 配比。 +3. **只看 loss 会漏掉工具调用、事实性、安全和交互能力。**：只看 loss 会漏掉工具调用、事实性、安全和交互能力。 +4. **外推不能替代中途 checkpoint 监控**：外推不能替代中途 checkpoint 监控，训练崩掉时曲线也会骗你。 + +## 学到什么 + +- 大模型路线先是预算工程，再是模型魔法。 +- 一条可外推曲线能把“信仰扩参”变成“可审计决策”。 +- 后续 Chinchilla、PaLM、LLaMA 都在回应这类 compute allocation 问题。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[gpt-3]]、[[chinchilla]]、[[llama]]、[[palm-2022]] + +## 关联 + +- [[gpt-3]] +- [[chinchilla]] +- [[llama]] +- [[palm-2022]] + +## 反向链接 + + diff --git a/src/content/docs/papers/lamda-2022.md b/src/content/docs/papers/lamda-2022.md new file mode 100644 index 000000000..ca4c9ff1f --- /dev/null +++ b/src/content/docs/papers/lamda-2022.md @@ -0,0 +1,90 @@ +--- +title: 'LaMDA — 聊天模型先学会有用、具体和不乱说' +description: '用 LaMDA 理解开放域对话模型为什么需要质量、安全和 groundedness 三条线。' +来源: 'Thoppilan et al., arXiv:2201.08239' +日期: 2026-07-14 +分类: LLM / Dialogue +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2201.08239v3 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2201.08239 + source_version: arXiv:2201.08239v3 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v3 +--- + +## 是什么 + +LaMDA: Language Models for Dialog Applications 是一篇 LLM / Dialogue 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像训练客服新人：不只是能接话，还要回答具体、别冒犯人、遇到事实问题要查依据。 + +它在本轮 40 篇里的位置是 **Batch 2 / open and dialogue models**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +开放域聊天很容易变成“流畅废话”：模型能接上上下文，却可能空泛、危险或编造事实。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| 对话质量指标 | 用 sensibleness、specificity、interestingness 衡量聊天是否像样。 | +| 安全过滤与标注 | 把不安全回复作为独立目标处理。 | +| Groundedness | 对事实问题引入外部检索和引用意识。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +用户问“明天东京天气如何”，只靠模型参数回答就可能乱编；LaMDA 式流程会先判断这是事实查询，再引入外部来源，而不是把聊天流畅度当事实性。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **“有趣”可能和“安全”冲突**：“有趣”可能和“安全”冲突，不能只优化用户停留时长。 +2. **groundedness 不是简单贴链接**：groundedness 不是简单贴链接，链接必须支撑回答里的具体断言。 +3. **安全分类器会有文化和语言边界**：安全分类器会有文化和语言边界，不能当一次性解决方案。 +4. **开放域对话的评测高度依赖人工偏好**：开放域对话的评测高度依赖人工偏好，自动分数只能辅助。 + +## 学到什么 + +- 聊天模型是质量、安全、事实性的多目标优化。 +- LaMDA 把对话产品的验收从“像人说话”推进到“能安全服务”。 +- 后来的 Bard、Gemini 和 ChatGPT 评测都继承了这类问题拆分。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[gpt-3]]、[[webgpt-2021]]、[[truthfulqa-2021]]、[[constitutional-ai]] + +## 关联 + +- [[gpt-3]] +- [[webgpt-2021]] +- [[truthfulqa-2021]] +- [[constitutional-ai]] + +## 反向链接 + + diff --git a/src/content/docs/papers/least-to-most-prompting-2022.md b/src/content/docs/papers/least-to-most-prompting-2022.md new file mode 100644 index 000000000..d966cf2f1 --- /dev/null +++ b/src/content/docs/papers/least-to-most-prompting-2022.md @@ -0,0 +1,90 @@ +--- +title: 'Least-to-Most — 先拆小题，再解大题' +description: '用 Least-to-Most Prompting 理解复杂推理为什么要先分解再逐步求解。' +来源: 'Zhou et al., arXiv:2205.10625' +日期: 2026-07-14 +分类: LLM / Reasoning +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2205.10625v3 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2205.10625 + source_version: arXiv:2205.10625v3 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v3 +--- + +## 是什么 + +Least-to-Most Prompting Enables Complex Reasoning in Large Language Models 是一篇 LLM / Reasoning 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像解奥数题先问“这题能拆成哪几个小问”，而不是一口气从题干跳到答案。 + +它在本轮 40 篇里的位置是 **Batch 4 / reasoning prompts**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +Chain-of-Thought 能让模型写步骤，但面对组合性强的问题，模型仍可能在第一步就选错路线。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| 问题分解 | 先生成一串更小、更容易的问题。 | +| 逐步求解 | 每个小问题用前面答案作为上下文。 | +| 组合泛化 | 测试模型能否把学过的小技能组合到更长任务。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +问“如果 Alice 比 Bob 多 3 个苹果，Bob 又比 Carol 多 2 个，Carol 有 5 个，Alice 有几个？”先拆 Carol->Bob，再 Bob->Alice，错误比直接心算少。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **第一步分解错**：第一步分解错，后面会稳定地错下去。 +2. **并非所有任务都适合线性拆解**：并非所有任务都适合线性拆解，有些需要回溯。 +3. **prompt 更长会增加成本和上下文噪声。**：prompt 更长会增加成本和上下文噪声。 +4. **分解质量需要单独评估**：分解质量需要单独评估，不能只看最终答案。 + +## 学到什么 + +- 复杂推理常常先是任务编排问题，再是单步能力问题。 +- Least-to-Most 是 agent planner 的早期 prompt 形态。 +- 后续 Plan-and-Solve、Tree of Thoughts 都在扩展这条线。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[chain-of-thought]]、[[tree-of-thoughts-2023]]、[[plan-and-solve-prompting-2023]]、[[self-consistency-2022]] + +## 关联 + +- [[chain-of-thought]] +- [[tree-of-thoughts-2023]] +- [[plan-and-solve-prompting-2023]] +- [[self-consistency-2022]] + +## 反向链接 + + diff --git a/src/content/docs/papers/linformer-2020.md b/src/content/docs/papers/linformer-2020.md new file mode 100644 index 000000000..421938512 --- /dev/null +++ b/src/content/docs/papers/linformer-2020.md @@ -0,0 +1,90 @@ +--- +title: 'Linformer — 把 attention 矩阵投影成线性复杂度' +description: '用 Linformer 理解低秩假设如何压缩 self-attention。' +来源: 'Wang et al., arXiv:2006.04768' +日期: 2026-07-14 +分类: NLP / Efficient Attention +难度: 高级 +difficulty: advanced +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2006.04768v3 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2006.04768 + source_version: arXiv:2006.04768v3 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v3 +--- + +## 是什么 + +Linformer: Self-Attention with Linear Complexity 是一篇 Transformer / Efficient Attention 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像把一张超大照片先压缩成少数关键列，再做分析；不是每个像素都两两比较。 + +它在本轮 40 篇里的位置是 **Batch 8 / long context and inference**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +标准 self-attention 对序列长度是 O(n²)，长文本会迅速耗尽显存和计算。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| 低秩假设 | 认为 attention 矩阵可以用较低维度近似。 | +| K/V 投影 | 把 key 和 value 沿序列维压到固定 k。 | +| 线性复杂度 | 把长序列成本从平方级降到近似线性。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +10000 个 token 两两 attention 要 1 亿级关系；Linformer 先投影到 256 个摘要位置，再计算关系。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **低秩近似不是所有任务都成立**：低秩近似不是所有任务都成立，细粒度长程依赖可能受损。 +2. **固定投影维度需要按任务和长度调。**：固定投影维度需要按任务和长度调。 +3. **它解决的是 attention 成本**：它解决的是 attention 成本，不解决所有长上下文记忆问题。 +4. **后续 Performer、Longformer、BigBird 走了不同取舍。**：后续 Performer、Longformer、BigBird 走了不同取舍。 + +## 学到什么 + +- 高效 attention 的本质是承认不是每个 token 对都同等重要。 +- Linformer 是长上下文效率路线的重要早期方案。 +- 产品选型要看任务依赖模式，而不是只看复杂度公式。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[performer-2020]]、[[longformer-2020]]、[[bigbird-2020]]、[[reformer-2020]] + +## 关联 + +- [[performer-2020]] +- [[longformer-2020]] +- [[bigbird-2020]] +- [[reformer-2020]] + +## 反向链接 + + diff --git a/src/content/docs/papers/longnet-2023.md b/src/content/docs/papers/longnet-2023.md new file mode 100644 index 000000000..b155bbd29 --- /dev/null +++ b/src/content/docs/papers/longnet-2023.md @@ -0,0 +1,90 @@ +--- +title: 'LongNet — 用 dilated attention 把上下文推到十亿 token 想象空间' +description: '用 LongNet 理解扩张式 attention 如何在多尺度上连接超长序列。' +来源: 'Ding et al., arXiv:2307.02486' +日期: 2026-07-14 +分类: NLP / Long Context +难度: 高级 +difficulty: advanced +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2307.02486v2 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2307.02486 + source_version: arXiv:2307.02486v2 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v2 +--- + +## 是什么 + +LongNet: Scaling Transformers to 1,000,000,000 Tokens 是一篇 Transformer / Long Context 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像城市交通：近处走小路，远处走高速，不需要每两个地点都修直达路。 + +它在本轮 40 篇里的位置是 **Batch 8 / long context and inference**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +百万级甚至更长上下文不能靠标准 attention 硬算，需要多尺度连接结构。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| Dilated attention | 按距离扩大 attention 间隔，覆盖更远范围。 | +| 分段设计 | 在局部细看、远处粗看之间折中。 | +| 超长上下文实验 | 验证在长序列建模中的扩展潜力。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +读 1000 页书时，当前页逐句看，上一章按段看，整本书按章节摘要看，这就是多尺度注意力直觉。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **超长上下文可训练不等于模型会有效使用全部信息。**：超长上下文可训练不等于模型会有效使用全部信息。 +2. **远距离粗看可能漏细节。**：远距离粗看可能漏细节。 +3. **benchmark 长度和真实任务长度不是一回事。**：benchmark 长度和真实任务长度不是一回事。 +4. **工程上还要解决数据加载、位置编码和推理内存。**：工程上还要解决数据加载、位置编码和推理内存。 + +## 学到什么 + +- LongNet 把长上下文问题从窗口扩展推进到多尺度结构。 +- 长上下文能力需要“能放进去”和“能找出来”两套评测。 +- 它适合和 RAG、记忆系统一起比较，而不是互相替代。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[bigbird-2020]]、[[longformer-2020]]、[[gemini-1.5-2024]]、[[memgym]] + +## 关联 + +- [[bigbird-2020]] +- [[longformer-2020]] +- [[gemini-1.5-2024]] +- [[memgym]] + +## 反向链接 + + diff --git a/src/content/docs/papers/minerva-2022.md b/src/content/docs/papers/minerva-2022.md new file mode 100644 index 000000000..171a095c4 --- /dev/null +++ b/src/content/docs/papers/minerva-2022.md @@ -0,0 +1,90 @@ +--- +title: 'Minerva — 把语言模型拉进数学草稿纸' +description: '用 Minerva 理解为什么数学推理需要专门的数据、逐步解题和采样验证。' +来源: 'Lewkowycz et al., arXiv:2206.14858' +日期: 2026-07-14 +分类: LLM / Math Reasoning +难度: 高级 +difficulty: advanced +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2206.14858v2 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2206.14858 + source_version: arXiv:2206.14858v2 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v2 +--- + +## 是什么 + +Solving Quantitative Reasoning Problems with Language Models 是一篇 LLM / Math Reasoning 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像让一个语文很强的学生转去参加数学竞赛：会读题还不够，还要见过足够多推导格式，并愿意把草稿一步步写出来。 + +它在本轮 40 篇里的位置是 **Batch 1 / foundation scaling**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +通用 LLM 在自然语言上强，但数学和科学题需要符号、公式、长链推理和计算一致性，普通网页语料不足以稳定支持。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| 数学/科学语料继续训练 | 让模型多见 LaTeX、公式和定量推导。 | +| 逐步解题格式 | 鼓励模型输出中间步骤，而不是只给最终答案。 | +| 采样与投票 | 多次生成候选解，用一致性提高最终准确率。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +同一道应用题让模型只报答案，很容易算错；让它列方程、化简、再代入，并采样 5 次取多数，错误率会下降。这就是 Minerva 的产品直觉。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **采样投票提高的是答案选择**：采样投票提高的是答案选择，不保证每条推导都严谨。 +2. **数学语料会带来格式优势**：数学语料会带来格式优势，但不能替代形式化证明器。 +3. **模型可能写出看似漂亮但中间偷换概念的推导。**：模型可能写出看似漂亮但中间偷换概念的推导。 +4. **只看竞赛题会高估真实工程计算和数据分析能力。**：只看竞赛题会高估真实工程计算和数据分析能力。 + +## 学到什么 + +- 数学能力是“数据分布 + 推理格式 + 验证策略”的组合。 +- Minerva 是后续 GSM8K、PAL、Program of Thoughts 的重要背景。 +- 对产品来说，长推理必须配合检查器或工具，而不是只相信自然语言草稿。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[gsm8k-2021]]、[[program-of-thoughts-2022]]、[[pal-code-reasoning-2022]]、[[self-consistency-2022]] + +## 关联 + +- [[gsm8k-2021]] +- [[program-of-thoughts-2022]] +- [[pal-code-reasoning-2022]] +- [[self-consistency-2022]] + +## 反向链接 + + diff --git a/src/content/docs/papers/mistral-7b-2023.md b/src/content/docs/papers/mistral-7b-2023.md new file mode 100644 index 000000000..ae9ca735e --- /dev/null +++ b/src/content/docs/papers/mistral-7b-2023.md @@ -0,0 +1,90 @@ +--- +title: 'Mistral 7B — 小模型靠架构细节打出性价比' +description: '用 Mistral 7B 理解 grouped-query attention 和 sliding-window attention 如何服务高效开源模型。' +来源: 'Jiang et al., arXiv:2310.06825' +日期: 2026-07-14 +分类: LLM / Efficient Model +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2310.06825v1 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2310.06825 + source_version: arXiv:2310.06825v1 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v1 +--- + +## 是什么 + +Mistral 7B 是一篇 LLM / Efficient Model 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像一辆轻量赛车：马力不是最大，但换挡、风阻和轮胎都调得很准，所以单位成本跑得快。 + +它在本轮 40 篇里的位置是 **Batch 2 / open and dialogue models**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +开源社区需要能本地部署和微调的强模型，不能每个场景都依赖 70B 或更大的底座。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| Grouped-Query Attention | 减少 KV cache 成本，让推理更省。 | +| Sliding-Window Attention | 让模型关注局部窗口，控制长序列计算。 | +| 强基线评测 | 用 7B 规模挑战更大模型的通用能力。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +把客服 FAQ 部署在单卡机器上，70B 模型延迟和显存都吃紧；7B 模型如果结构高效，可以在可接受质量下显著降低服务成本。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **小模型强不等于所有任务都够用**：小模型强不等于所有任务都够用，复杂推理仍可能需要更大模型或工具。 +2. **sliding window 会改变长程依赖处理方式**：sliding window 会改变长程依赖处理方式，长文任务要单独验收。 +3. **benchmark 优势不能直接等于业务指标**：benchmark 优势不能直接等于业务指标，需要按场景复测。 +4. **开源权重仍要看许可证和商用限制。**：开源权重仍要看许可证和商用限制。 + +## 学到什么 + +- 模型效率来自架构、训练数据和部署约束的共同设计。 +- Mistral 7B 是“够强且够便宜”的开源模型代表。 +- 它让产品团队更容易把 LLM 带到私有化和边缘场景。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[llama]]、[[paged-attention]]、[[qlora-2023]]、[[speculative-decoding-2022]] + +## 关联 + +- [[llama]] +- [[paged-attention]] +- [[qlora-2023]] +- [[speculative-decoding-2022]] + +## 反向链接 + + diff --git a/src/content/docs/papers/mrkl-systems-2022.md b/src/content/docs/papers/mrkl-systems-2022.md new file mode 100644 index 000000000..89ef5ed81 --- /dev/null +++ b/src/content/docs/papers/mrkl-systems-2022.md @@ -0,0 +1,90 @@ +--- +title: 'MRKL — 给大模型配一组专家工具和路由器' +description: '用 MRKL Systems 理解 neuro-symbolic agent 为什么要把 LLM、检索和计算模块拆开。' +来源: 'Karpas et al., arXiv:2205.00445' +日期: 2026-07-14 +分类: LLM / Tool Architecture +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2205.00445v1 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2205.00445 + source_version: arXiv:2205.00445v1 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v1 +--- + +## 是什么 + +MRKL Systems: A modular, neuro-symbolic architecture that combines large language models, external knowledge sources and discrete reasoning 是一篇 LLM / Tool Architecture 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像医院分诊台：LLM 不必自己做所有检查，而是判断该去影像科、化验科还是专家门诊。 + +它在本轮 40 篇里的位置是 **Batch 5 / agents and tools**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +单个语言模型既要懂语言、查事实、算数、调用业务系统，可靠性和可维护性都会变差。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| 模块化专家 | 把计算器、搜索、数据库、规则系统等做成独立 expert。 | +| 路由/编排 | 由模型或控制器决定什么时候调用哪个模块。 | +| 神经 + 符号结合 | 让 LLM 做语言理解，让确定性系统做可验证操作。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +用户问“把 17 美元按今天汇率换成人民币再加 6% 税是多少”，MRKL 会路由到汇率工具和计算器，而不是让 LLM 心算。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **路由错了**：路由错了，比不用工具更糟。 +2. **专家模块接口要稳定**：专家模块接口要稳定，否则 prompt 里写得再好也会失败。 +3. **工具结果需要回填上下文**：工具结果需要回填上下文，避免模型忽略真实返回。 +4. **模块越多**：模块越多，权限和审计越重要。 + +## 学到什么 + +- MRKL 把 agent 可靠性问题转成系统架构问题。 +- 今天的 function calling、MCP 和 tool router 都能看到它的影子。 +- LLM 产品不该追求一个模型包打天下。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[toolformer]]、[[mcp-bench-2025]]、[[program-of-thoughts-2022]]、[[gorilla-2023]] + +## 关联 + +- [[toolformer]] +- [[mcp-bench-2025]] +- [[program-of-thoughts-2022]] +- [[gorilla-2023]] + +## 反向链接 + + diff --git a/src/content/docs/papers/natural-instructions-v2-2022.md b/src/content/docs/papers/natural-instructions-v2-2022.md new file mode 100644 index 000000000..18eac7ea0 --- /dev/null +++ b/src/content/docs/papers/natural-instructions-v2-2022.md @@ -0,0 +1,90 @@ +--- +title: 'Super-NaturalInstructions — 1600+ 任务教模型读懂说明书' +description: '用 Super-NaturalInstructions 理解 declarative instructions 如何评测任务泛化。' +来源: 'Wang et al., arXiv:2204.07705' +日期: 2026-07-14 +分类: LLM / Instruction Benchmark +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2204.07705v3 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2204.07705 + source_version: arXiv:2204.07705v3 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v3 +--- + +## 是什么 + +Super-NaturalInstructions: Generalization via Declarative Instructions on 1600+ NLP Tasks 是一篇 LLM / Instruction Benchmark 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像给实习生一本任务说明书：没见过这个任务也要靠说明、正反例和约束完成。 + +它在本轮 40 篇里的位置是 **Batch 3 / instruction tuning**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +传统 NLP benchmark 常按任务训练/测试，难以回答“模型能否读懂新任务说明并迁移”。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| 1600+ 任务集合 | 覆盖分类、生成、改写等大量 NLP 任务。 | +| Declarative instruction | 用自然语言写清任务定义和输出要求。 | +| 跨任务泛化评测 | 训练和测试任务分开，检查新任务适应能力。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +模型从没见过“把评论改写成更礼貌语气”，但说明书给了定义和例子。它如果能完成，说明学到的是读说明执行，而不是背任务标签。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **任务说明质量会强烈影响表现**：任务说明质量会强烈影响表现，说明写得差会误伤模型。 +2. **任务多不等于真实场景全覆盖**：任务多不等于真实场景全覆盖，长程工具任务仍缺失。 +3. **模型可能依赖表面关键词**：模型可能依赖表面关键词，而不是真正理解任务定义。 +4. **多任务数据集需要严防训练/测试污染。**：多任务数据集需要严防训练/测试污染。 + +## 学到什么 + +- Instruction following 可以被拆成“读说明、看例子、执行约束”。 +- 它为 Self-Instruct、FLAN、WizardLM 提供了任务泛化背景。 +- 做 agent 评测时，也应该把任务说明质量纳入变量。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[flan-2021]]、[[self-instruct-2022]]、[[wizardlm-2023]]、[[t5]] + +## 关联 + +- [[flan-2021]] +- [[self-instruct-2022]] +- [[wizardlm-2023]] +- [[t5]] + +## 反向链接 + + diff --git a/src/content/docs/papers/opt-2022.md b/src/content/docs/papers/opt-2022.md new file mode 100644 index 000000000..22a70c875 --- /dev/null +++ b/src/content/docs/papers/opt-2022.md @@ -0,0 +1,90 @@ +--- +title: 'OPT — 把 GPT-3 级训练日志打开给研究社区' +description: '用 OPT 理解开放权重、训练日志和复现实验对 LLM 研究的重要性。' +来源: 'Zhang et al., arXiv:2205.01068' +日期: 2026-07-14 +分类: LLM / Open Model +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2205.01068v4 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2205.01068 + source_version: arXiv:2205.01068v4 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v4 +--- + +## 是什么 + +OPT: Open Pre-trained Transformer Language Models 是一篇 LLM / Open Model 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像不仅把成品菜端出来，还把采购单、翻车记录和厨房温度一起公开，别人才能判断这道菜怎么做出来的。 + +它在本轮 40 篇里的位置是 **Batch 2 / open and dialogue models**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +GPT-3 证明了少样本能力，但闭源让学术界难以复查训练细节、失败模式和安全风险。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| 多尺度开放模型 | 发布从小到 175B 的 OPT 系列。 | +| 训练日志披露 | 记录训练中断、数据和工程问题。 | +| 研究访问机制 | 让更多研究者能在同一底座上分析行为。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +如果两个团队都声称“复现 GPT-3”，但一个只给结果，一个给 checkpoint 和训练日志，后者才能让你定位损失尖峰、数据污染和评测差异。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **开放权重不等于训练完全可复现**：开放权重不等于训练完全可复现，算力和数据仍是门槛。 +2. **开放模型也会传播滥用风险**：开放模型也会传播滥用风险，需要访问策略和用途边界。 +3. **OPT 的能力不是最新最强**：OPT 的能力不是最新最强，但透明度本身是贡献。 +4. **训练日志是研究资产**：训练日志是研究资产，不应只在失败时才记录。 + +## 学到什么 + +- LLM 研究需要可检查的中间过程，而不只是 leaderboard。 +- OPT 是开放模型治理和复现文化的重要节点。 +- 后续 BLOOM、LLaMA、Mistral 都在不同程度回应开放性问题。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[gpt-3]]、[[bloom-2022]]、[[llama]]、[[mistral-7b-2023]] + +## 关联 + +- [[gpt-3]] +- [[bloom-2022]] +- [[llama]] +- [[mistral-7b-2023]] + +## 反向链接 + + diff --git a/src/content/docs/papers/orca-explanation-tuning-2023.md b/src/content/docs/papers/orca-explanation-tuning-2023.md new file mode 100644 index 000000000..fcfacaacc --- /dev/null +++ b/src/content/docs/papers/orca-explanation-tuning-2023.md @@ -0,0 +1,90 @@ +--- +title: 'Orca — 小模型不只抄答案，还学解释轨迹' +description: '用 Orca 理解 explanation tuning 为什么比只蒸馏最终答案更像教学生。' +来源: 'Mukherjee et al., arXiv:2306.02707' +日期: 2026-07-14 +分类: LLM / Distillation +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2306.02707v1 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2306.02707 + source_version: arXiv:2306.02707v1 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v1 +--- + +## 是什么 + +Orca: Progressive Learning from Complex Explanation Traces of GPT-4 是一篇 LLM / Distillation 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像学霸讲题不是只给答案 C，而是把为什么排除 A/B/D 的过程写出来，普通学生才更容易迁移。 + +它在本轮 40 篇里的位置是 **Batch 3 / instruction tuning**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +传统蒸馏常让小模型模仿大模型答案，但复杂任务真正有价值的是解题步骤、解释和中间决策。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| Explanation traces | 收集 GPT-4 等教师模型的详细解释。 | +| Progressive learning | 从简单到复杂组织训练信号。 | +| 多任务蒸馏 | 覆盖推理、写作和理解任务，而不局限单一 benchmark。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +同一道逻辑题，小模型只看“答案是 7”学不到方法；看“先列变量、再代入、最后检查约束”才可能迁移到新题。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **教师解释可能也有错**：教师解释可能也有错，长解释不自动等于真推理。 +2. **蒸馏会继承教师风格和偏见。**：蒸馏会继承教师风格和偏见。 +3. **如果评测题和教师数据太近**：如果评测题和教师数据太近，泛化会被高估。 +4. **小模型学会解释口吻**：小模型学会解释口吻，不代表内部机制真的等同教师。 + +## 学到什么 + +- Orca 把蒸馏目标从答案推进到过程。 +- 这对企业小模型很重要：预算有限时，可以买教师轨迹而不是只买标签。 +- 解释轨迹仍需事实校验和任务外验证。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[ccopd-distillation]]、[[self-instruct-2022]]、[[wizardlm-2023]]、[[gsm8k-2021]] + +## 关联 + +- [[ccopd-distillation]] +- [[self-instruct-2022]] +- [[wizardlm-2023]] +- [[gsm8k-2021]] + +## 反向链接 + + diff --git a/src/content/docs/papers/p-tuning-v2-2021.md b/src/content/docs/papers/p-tuning-v2-2021.md new file mode 100644 index 000000000..69cb215b0 --- /dev/null +++ b/src/content/docs/papers/p-tuning-v2-2021.md @@ -0,0 +1,90 @@ +--- +title: 'P-Tuning v2 — 把 prompt tuning 深插到每一层' +description: '用 P-Tuning v2 理解深层连续提示为什么能跨规模和任务接近 fine-tuning。' +来源: 'Liu et al., arXiv:2110.07602' +日期: 2026-07-14 +分类: LLM / Efficient Finetuning +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2110.07602v3 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2110.07602 + source_version: arXiv:2110.07602v3 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v3 +--- + +## 是什么 + +P-Tuning v2: Prompt Tuning Can Be Comparable to Fine-tuning Universally Across Scales and Tasks 是一篇 LLM / Efficient Finetuning 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像不只在书的第一页贴提示，而是在每一章开头都放一张任务提醒。 + +它在本轮 40 篇里的位置是 **Batch 7 / parameter-efficient tuning**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +早期 prompt tuning 在小模型和复杂序列标注任务上不稳定，离 full fine-tuning 还有差距。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| Deep prompt | 在多层加入可训练提示，而不是只改输入层。 | +| 跨任务测试 | 覆盖 NLU、序列标注等更多任务。 | +| 冻结主干 | 仍保留参数高效和多任务存储优势。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +做命名实体识别时，只在输入前加 soft prompt 不够；深层 prompt 可以在模型内部多处影响表示。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **插入层数越多**：插入层数越多，工程实现越依赖模型结构。 +2. **深层 prompt 更强**：深层 prompt 更强，但也更难解释。 +3. **不同任务的 prompt 长度和位置需要调参。**：不同任务的 prompt 长度和位置需要调参。 +4. **接近 fine-tuning 不代表所有分布外场景都稳。**：接近 fine-tuning 不代表所有分布外场景都稳。 + +## 学到什么 + +- PEFT 的设计空间不止“训练多少参数”，还有“参数插在哪里”。 +- P-Tuning v2 把 prompt tuning 从生成任务推向更通用 NLU。 +- 它提醒我们：轻量适配也需要体系化工程。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[prompt-tuning-2021]]、[[prefix-tuning-2021]]、[[lora]]、[[qlora-2023]] + +## 关联 + +- [[prompt-tuning-2021]] +- [[prefix-tuning-2021]] +- [[lora]] +- [[qlora-2023]] + +## 反向链接 + + diff --git a/src/content/docs/papers/pal-code-reasoning-2022.md b/src/content/docs/papers/pal-code-reasoning-2022.md new file mode 100644 index 000000000..9eb215a36 --- /dev/null +++ b/src/content/docs/papers/pal-code-reasoning-2022.md @@ -0,0 +1,90 @@ +--- +title: 'PAL — 让 Python 成为语言模型的草稿纸' +description: '用 PAL 理解 Program-aided Language Models 如何把推理转成可运行代码。' +来源: 'Gao et al., arXiv:2211.10435' +日期: 2026-07-14 +分类: LLM / Tool Reasoning +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2211.10435v2 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2211.10435 + source_version: arXiv:2211.10435v2 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v2 +--- + +## 是什么 + +PAL: Program-aided Language Models 是一篇 LLM / Tool Reasoning 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像数学老师要求你别只写“显然”，而是写一段能跑的 Python 来证明答案。 + +它在本轮 40 篇里的位置是 **Batch 4 / reasoning prompts**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +自然语言 chain-of-thought 在算术、日期、组合题上容易出现局部算错，但模型又能写出接近正确的程序结构。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| 代码形式推理 | 把中间步骤表达为 Python 程序。 | +| 外部解释器 | 用真实执行结果作为答案。 | +| 少样本提示 | 通过示例教模型生成合适代码。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +问“第 100 个偶数是多少”，模型生成 `2 * 100` 比写一段自然语言解释更不容易漂移。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **生成代码可能通过测试但语义不对**：生成代码可能通过测试但语义不对，需要边界样例。 +2. **解释器让答案确定**：解释器让答案确定，但不保证题意理解正确。 +3. **代码工具对非程序员用户不可见**：代码工具对非程序员用户不可见，产品要把结果解释回自然语言。 +4. **沙箱权限、超时和依赖管理是工程必备项。**：沙箱权限、超时和依赖管理是工程必备项。 + +## 学到什么 + +- PAL 把“会推理”转成“能执行”，这是 agent 产品的关键转变。 +- 工具调用不是 LLM 的外挂，而是可靠性结构。 +- 它和 Program of Thoughts 共同奠定了代码执行式推理路线。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[program-of-thoughts-2022]]、[[codex-2021]]、[[toolformer]]、[[swe-bench]] + +## 关联 + +- [[program-of-thoughts-2022]] +- [[codex-2021]] +- [[toolformer]] +- [[swe-bench]] + +## 反向链接 + + diff --git a/src/content/docs/papers/palm-2022.md b/src/content/docs/papers/palm-2022.md new file mode 100644 index 000000000..46309124d --- /dev/null +++ b/src/content/docs/papers/palm-2022.md @@ -0,0 +1,90 @@ +--- +title: 'PaLM — Pathways 把 540B LLM 扩成统一底座' +description: '用 PaLM 理解大规模 dense decoder 如何在多任务、推理和代码能力上同时冒头。' +来源: 'Chowdhery et al., arXiv:2204.02311' +日期: 2026-07-14 +分类: LLM / Foundation Model +难度: 高级 +difficulty: advanced +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2204.02311v5 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2204.02311 + source_version: arXiv:2204.02311v5 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v5 +--- + +## 是什么 + +PaLM: Scaling Language Modeling with Pathways 是一篇 LLM / Foundation Model 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像把很多单科老师请进一间超级教室：模型本体还是同一个 decoder，但它在语言、数学、代码、多语种任务上都开始表现出可迁移能力。 + +它在本轮 40 篇里的位置是 **Batch 1 / foundation scaling**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +GPT-3 之后，行业想知道继续扩 dense decoder 是否还能带来跨任务收益，以及工程系统能否支撑 500B 级训练。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| Pathways 训练栈 | 用 Google 的分布式系统承载 540B 参数训练。 | +| 统一 decoder-only LM | 不为每个任务改架构，而是靠规模和数据覆盖。 | +| 大规模评测 | 同时看语言理解、推理、代码和多语种能力。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +把一个小模型分别拿去做翻译、数学和代码补全，会看到能力割裂；PaLM 的核心观察是同一底座规模变大后，少样本提示能跨更多任务工作。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **PaLM 不是“只要变大就自动安全”**：PaLM 不是“只要变大就自动安全”，安全和事实性仍要单独治理。 +2. **论文报告的是系统级结果**：论文报告的是系统级结果，普通团队不能按同样算力复现。 +3. **emergent ability 的表述容易被过度神化**：emergent ability 的表述容易被过度神化，很多能力依赖评测阈值和 prompting。 +4. **dense 540B 的服务成本很高**：dense 540B 的服务成本很高，后来 MoE、小模型和蒸馏都在补这个问题。 + +## 学到什么 + +- PaLM 证明 foundation model 可以成为多任务产品底座。 +- 规模收益必须和训练系统、数据治理、评测矩阵一起看。 +- 它是 Minerva、Flan-PaLM 等后续路线的重要上游。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[gpt-3]]、[[minerva-2022]]、[[ul2-2022]]、[[chinchilla]] + +## 关联 + +- [[gpt-3]] +- [[minerva-2022]] +- [[ul2-2022]] +- [[chinchilla]] + +## 反向链接 + + diff --git a/src/content/docs/papers/plan-and-solve-prompting-2023.md b/src/content/docs/papers/plan-and-solve-prompting-2023.md new file mode 100644 index 000000000..43f2b427e --- /dev/null +++ b/src/content/docs/papers/plan-and-solve-prompting-2023.md @@ -0,0 +1,90 @@ +--- +title: 'Plan-and-Solve — 零样本推理先写计划再执行' +description: '用 Plan-and-Solve 理解为什么 prompt 可以显式拆成 plan 和 solve 两段。' +来源: 'Wang et al., arXiv:2305.04091' +日期: 2026-07-14 +分类: LLM / Reasoning +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2305.04091v3 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2305.04091 + source_version: arXiv:2305.04091v3 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v3 +--- + +## 是什么 + +Plan-and-Solve Prompting: Improving Zero-Shot Chain-of-Thought Reasoning by Large Language Models 是一篇 LLM / Reasoning 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像考试前先列提纲：第一步读题，第二步列公式，第三步检查单位，然后才开始写答案。 + +它在本轮 40 篇里的位置是 **Batch 6 / agent tool ecosystems**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +Zero-shot CoT 只要求“Let’s think step by step”，但没有约束模型先形成完整计划，容易漏步骤。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| Plan phase | 让模型先写解决方案大纲。 | +| Solve phase | 按计划逐步执行并给答案。 | +| 增强版 prompt | 加入计算、遗漏检查等提示降低常见错误。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +处理报销题时，计划先列“汇总金额、扣除不可报销项、计算税费、输出结果”，再逐项求解，能减少漏扣项目。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **计划写得漂亮但不可执行时**：计划写得漂亮但不可执行时，solve 阶段仍会失败。 +2. **简单题强行计划会增加冗余。**：简单题强行计划会增加冗余。 +3. **prompt 结构收益依赖模型基础能力。**：prompt 结构收益依赖模型基础能力。 +4. **没有外部验证时**：没有外部验证时，计划和答案可能一起错。 + +## 学到什么 + +- Plan-and-Solve 是轻量 planner，不需要训练就能改善部分推理。 +- 它连接了 CoT prompt 和 agent planning。 +- 产品中可以把“计划可见化”作为用户信任和调试入口。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[least-to-most-prompting-2022]]、[[chain-of-thought]]、[[agent-planning-benchmark-2026]]、[[tree-of-thoughts-2023]] + +## 关联 + +- [[least-to-most-prompting-2022]] +- [[chain-of-thought]] +- [[agent-planning-benchmark-2026]] +- [[tree-of-thoughts-2023]] + +## 反向链接 + + diff --git a/src/content/docs/papers/prefix-tuning-2021.md b/src/content/docs/papers/prefix-tuning-2021.md new file mode 100644 index 000000000..be51c6eb5 --- /dev/null +++ b/src/content/docs/papers/prefix-tuning-2021.md @@ -0,0 +1,90 @@ +--- +title: 'Prefix-Tuning — 不改模型，只给每层塞一段可训练前缀' +description: '用 Prefix-Tuning 理解连续 prompt 如何成为参数高效微调方法。' +来源: 'Li and Liang, arXiv:2101.00190' +日期: 2026-07-14 +分类: LLM / Efficient Finetuning +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2101.00190v1 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2101.00190 + source_version: arXiv:2101.00190v1 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v1 +--- + +## 是什么 + +Prefix-Tuning: Optimizing Continuous Prompts for Generation 是一篇 LLM / Efficient Finetuning 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像不重写整本说明书，只在每章前面贴一张任务提示卡，让读者按新任务理解后文。 + +它在本轮 40 篇里的位置是 **Batch 7 / parameter-efficient tuning**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +全量 fine-tuning 每个任务都复制一套模型参数，存储和维护成本高。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| 冻结 LM | 底座模型参数不更新。 | +| 可训练 prefix | 在每层 attention 前加入连续向量。 | +| 任务专属小参数 | 每个任务只保存很小的 prefix。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +同一个摘要模型要适配“新闻摘要”和“法律摘要”。Prefix-Tuning 不复制整个模型，只保存两套小 prefix。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **prefix 是连续向量**：prefix 是连续向量，人类不可读，调试不如自然语言 prompt 直观。 +2. **任务差异很大时**：任务差异很大时，小 prefix 可能容量不够。 +3. **不同架构的 prefix 插入点不同**：不同架构的 prefix 插入点不同，迁移要重做。 +4. **服务多个 prefix 时要管理加载和缓存。**：服务多个 prefix 时要管理加载和缓存。 + +## 学到什么 + +- 参数高效微调的核心是冻结通用知识，只学习任务控制面。 +- Prefix-Tuning 是 LoRA、Prompt Tuning 等 PEFT 方法的重要前奏。 +- 产品上它对应“同一底座，多套轻量行为配置”。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[prompt-tuning-2021]]、[[p-tuning-v2-2021]]、[[lora]]、[[qlora-2023]] + +## 关联 + +- [[prompt-tuning-2021]] +- [[p-tuning-v2-2021]] +- [[lora]] +- [[qlora-2023]] + +## 反向链接 + + diff --git a/src/content/docs/papers/program-of-thoughts-2022.md b/src/content/docs/papers/program-of-thoughts-2022.md new file mode 100644 index 000000000..83abae18d --- /dev/null +++ b/src/content/docs/papers/program-of-thoughts-2022.md @@ -0,0 +1,90 @@ +--- +title: 'Program of Thoughts — 让模型写程序，把计算交给解释器' +description: '用 Program of Thoughts 理解自然语言推理和精确计算为什么要分工。' +来源: 'Chen et al., arXiv:2211.12588' +日期: 2026-07-14 +分类: LLM / Tool Reasoning +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2211.12588v4 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2211.12588 + source_version: arXiv:2211.12588v4 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v4 +--- + +## 是什么 + +Program of Thoughts Prompting: Disentangling Computation from Reasoning for Numerical Reasoning Tasks 是一篇 LLM / Tool Reasoning 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像人做应用题：读题和列式靠脑子，真正大数计算交给计算器。 + +它在本轮 40 篇里的位置是 **Batch 4 / reasoning prompts**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +LLM 很会解释，但在多位数运算、循环和表格计算上容易算错。自然语言步骤不适合承担精确执行。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| 生成程序 | 让模型把题目转成可执行代码。 | +| 解释器执行 | 把算术和循环交给 Python 等工具。 | +| 推理/计算解耦 | 模型负责建模，工具负责确定性结果。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +题目要求算 37 个商品每个 19.8 元再打 85 折。模型写 `37 * 19.8 * 0.85`，解释器给出数值，避免口算漂移。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **程序写错比算错更隐蔽**：程序写错比算错更隐蔽，需要测试输入或断言。 +2. **代码执行有安全边界**：代码执行有安全边界，不能随便跑不可信代码。 +3. **有些题的难点是建模**：有些题的难点是建模，不是计算；解释器救不了错误公式。 +4. **工具调用延迟和沙箱成本要纳入产品设计。**：工具调用延迟和沙箱成本要纳入产品设计。 + +## 学到什么 + +- LLM 工具使用的第一原则是让确定性系统做确定性工作。 +- Program of Thoughts 是 PAL、Toolformer、agent tool use 的重要前身。 +- 代码不是装饰，而是把推理结果变成可执行证据。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[pal-code-reasoning-2022]]、[[toolformer]]、[[gsm8k-2021]]、[[react-agent]] + +## 关联 + +- [[pal-code-reasoning-2022]] +- [[toolformer]] +- [[gsm8k-2021]] +- [[react-agent]] + +## 反向链接 + + diff --git a/src/content/docs/papers/prompt-to-prompt-2022.md b/src/content/docs/papers/prompt-to-prompt-2022.md new file mode 100644 index 000000000..e83331cca --- /dev/null +++ b/src/content/docs/papers/prompt-to-prompt-2022.md @@ -0,0 +1,90 @@ +--- +title: 'Prompt-to-Prompt — 改词不改构图的 cross-attention 编辑' +description: '用 Prompt-to-Prompt 理解扩散模型里文本 token 和图像布局如何对齐。' +来源: 'Hertz et al., arXiv:2208.01626' +日期: 2026-07-14 +分类: Diffusion / Editing +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2208.01626v1 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2208.01626 + source_version: arXiv:2208.01626v1 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v1 +--- + +## 是什么 + +Prompt-to-Prompt Image Editing with Cross Attention Control 是一篇 Diffusion / Editing 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像在设计稿里把“红色汽车”改成“蓝色汽车”，但不希望车的位置、角度和背景全变。 + +它在本轮 40 篇里的位置是 **Batch 9 / controllable generation**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +扩散模型对 prompt 很敏感，小改一句话常导致整张图构图重排，编辑不可控。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| Cross-attention map | 利用文本 token 到图像区域的注意力关系。 | +| Attention 替换/冻结 | 编辑某些词时保留原布局注意力。 | +| 局部语义修改 | 让对象属性变，整体结构尽量不变。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +原 prompt“a cat sitting on a bench”，改成“a dog sitting on a bench”。保留 bench 和姿态 attention，只替换 cat/dog 相关区域。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **attention map 不是完美解释**：attention map 不是完美解释，复杂场景会错绑区域。 +2. **大幅语义改动无法保证构图不变。**：大幅语义改动无法保证构图不变。 +3. **方法依赖特定扩散采样和 cross-attention 结构。**：方法依赖特定扩散采样和 cross-attention 结构。 +4. **局部编辑仍可能产生边界伪影。**：局部编辑仍可能产生边界伪影。 + +## 学到什么 + +- Prompt-to-Prompt 把 prompt 编辑从随机试词推进到可控 attention 操作。 +- 生成模型的可编辑性来自中间表示，而不只是最终图片。 +- 设计工具需要“保留什么”和“改变什么”的显式控制面。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[controlnet-2023]]、[[textual-inversion-2022]]、[[ddim-2020]]、[[ddpm]] + +## 关联 + +- [[controlnet-2023]] +- [[textual-inversion-2022]] +- [[ddim-2020]] +- [[ddpm]] + +## 反向链接 + + diff --git a/src/content/docs/papers/prompt-tuning-2021.md b/src/content/docs/papers/prompt-tuning-2021.md new file mode 100644 index 000000000..0469bdd9d --- /dev/null +++ b/src/content/docs/papers/prompt-tuning-2021.md @@ -0,0 +1,90 @@ +--- +title: 'Prompt Tuning — 规模变大后，软提示也能接近微调' +description: '用 Prompt Tuning 理解为什么 soft prompt 在大模型上突然变得有效。' +来源: 'Lester et al., arXiv:2104.08691' +日期: 2026-07-14 +分类: LLM / Efficient Finetuning +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2104.08691v2 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2104.08691 + source_version: arXiv:2104.08691v2 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v2 +--- + +## 是什么 + +The Power of Scale for Parameter-Efficient Prompt Tuning 是一篇 LLM / Efficient Finetuning 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像给越聪明的学生越短的提示也够用：基础能力强了，小小提示就能调动已有知识。 + +它在本轮 40 篇里的位置是 **Batch 7 / parameter-efficient tuning**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +离散 prompt 需要人工写，full fine-tuning 又贵。问题是能不能只训练一小串连续 prompt token。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| Soft prompt | 在输入前加可训练 embedding。 | +| 冻结 T5 底座 | 只更新 prompt 参数。 | +| 规模效应分析 | 观察模型越大，prompt tuning 越接近 full tuning。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +把同一个 T5 用于情感分类，只训练 20 个虚拟 token，让它把任务映射到已有语言能力。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **小模型上 soft prompt 可能明显弱于 full fine-tuning。**：小模型上 soft prompt 可能明显弱于 full fine-tuning。 +2. **soft prompt 不可读**：soft prompt 不可读，不适合需要人工审查的控制策略。 +3. **prompt 长度、初始化和任务格式都影响结果。**：prompt 长度、初始化和任务格式都影响结果。 +4. **它主要调任务行为**：它主要调任务行为，不会凭空补充缺失知识。 + +## 学到什么 + +- 模型规模会改变适配方法的性价比。 +- Prompt Tuning 说明“控制模型”可以比“改模型”更轻。 +- PEFT 方法要结合底座规模一起评估。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[prefix-tuning-2021]]、[[p-tuning-v2-2021]]、[[t5]]、[[ul2-2022]] + +## 关联 + +- [[prefix-tuning-2021]] +- [[p-tuning-v2-2021]] +- [[t5]] +- [[ul2-2022]] + +## 反向链接 + + diff --git a/src/content/docs/papers/qlora-2023.md b/src/content/docs/papers/qlora-2023.md new file mode 100644 index 000000000..3f93fabd3 --- /dev/null +++ b/src/content/docs/papers/qlora-2023.md @@ -0,0 +1,90 @@ +--- +title: 'QLoRA — 4-bit 量化底座上贴 LoRA 也能微调' +description: '用 QLoRA 理解 NF4、double quantization 和 paged optimizers 如何降低微调门槛。' +来源: 'Dettmers et al., arXiv:2305.14314' +日期: 2026-07-14 +分类: LLM / Efficient Finetuning +难度: 高级 +difficulty: advanced +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2305.14314v1 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2305.14314 + source_version: arXiv:2305.14314v1 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v1 +--- + +## 是什么 + +QLoRA: Efficient Finetuning of Quantized LLMs 是一篇 LLM / Efficient Finetuning 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像把整套大机器冻在仓库里，只在外面接一小块可调控制板；机器本体还压缩到更省空间。 + +它在本轮 40 篇里的位置是 **Batch 7 / parameter-efficient tuning**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +LoRA 已经降低微调参数量，但大模型底座本身仍占显存，普通单卡很难微调 33B/65B 模型。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| NF4 量化 | 用适合正态权重的 4-bit 表示冻结底座。 | +| Double quantization | 继续压缩量化常数，节省显存。 | +| Paged optimizers | 用分页思想缓解 optimizer 显存峰值。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +一张消费级 GPU 放不下全精度 33B；QLoRA 把底座 4-bit 存放，只训练小 adapter，让个人实验成为可能。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **4-bit 微调不等于无损**：4-bit 微调不等于无损，极端任务仍要看质量回退。 +2. **显存省了**：显存省了，数据质量和评测仍是主要瓶颈。 +3. **adapter 合并、部署和多 adapter 管理会带来工程复杂度。**：adapter 合并、部署和多 adapter 管理会带来工程复杂度。 +4. **量化 kernel 和硬件支持会影响真实性能。**：量化 kernel 和硬件支持会影响真实性能。 + +## 学到什么 + +- QLoRA 把大模型微调从少数实验室推向普通团队。 +- 参数高效和内存高效要一起设计。 +- 它是开源 instruction model 爆发的重要基础设施。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[lora]]、[[mistral-7b-2023]]、[[deepspeed-zero]]、[[axolotl]] + +## 关联 + +- [[lora]] +- [[mistral-7b-2023]] +- [[deepspeed-zero]] +- [[axolotl]] + +## 反向链接 + + diff --git a/src/content/docs/papers/saycan-2022.md b/src/content/docs/papers/saycan-2022.md new file mode 100644 index 000000000..8d851f3e0 --- /dev/null +++ b/src/content/docs/papers/saycan-2022.md @@ -0,0 +1,90 @@ +--- +title: 'SayCan — 机器人不只问“想做什么”，还问“我能做什么”' +description: '用 SayCan 理解语言模型和机器人 affordance 如何合成可执行动作。' +来源: 'Ahn et al., arXiv:2204.01691' +日期: 2026-07-14 +分类: LLM Agent / Robotics +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2204.01691v2 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2204.01691 + source_version: arXiv:2204.01691v2 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v2 +--- + +## 是什么 + +Do As I Can, Not As I Say: Grounding Language in Robotic Affordances 是一篇 Embodied AI / Robotics 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像让人帮你做饭：他说“我应该切菜”是一回事，他手边有没有刀、会不会切、菜在不在台面上是另一回事。 + +它在本轮 40 篇里的位置是 **Batch 5 / agents and tools**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +LLM 能给出高层计划，但机器人必须选择当前环境里可执行、成功概率高的技能。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| 语言可用性 | LLM 评估某个技能是否符合用户指令。 | +| Affordance value | 机器人策略估计当前状态下技能能否成功。 | +| 乘积排序 | 把“应该做”和“做得到”合成动作选择。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +用户说“把饮料递给我”。LLM 觉得“拿起可乐”合理，但 affordance 发现可乐不在视野里、矿泉水在桌上，于是先选择可执行的抓取动作。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **LLM 计划正确但感知错**：LLM 计划正确但感知错，机器人仍会失败。 +2. **affordance 模型只覆盖已训练技能**：affordance 模型只覆盖已训练技能，超出技能库不能硬做。 +3. **乘积分数简单有效**：乘积分数简单有效，但复杂长程任务需要更强规划。 +4. **真实机器人安全约束不能只靠语言模型。**：真实机器人安全约束不能只靠语言模型。 + +## 学到什么 + +- 具身 agent 的关键是把语言意图接到可执行技能。 +- SayCan 是“LLM planner + skill library”路线的代表。 +- 软件 agent 也有类似问题：工具是否可用，比工具描述更重要。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[voyager]]、[[inner-monologue-2022]]、[[react-agent]]、[[osworld]] + +## 关联 + +- [[voyager]] +- [[inner-monologue-2022]] +- [[react-agent]] +- [[osworld]] + +## 反向链接 + + diff --git a/src/content/docs/papers/self-instruct-2022.md b/src/content/docs/papers/self-instruct-2022.md new file mode 100644 index 000000000..d873c8e91 --- /dev/null +++ b/src/content/docs/papers/self-instruct-2022.md @@ -0,0 +1,90 @@ +--- +title: 'Self-Instruct — 让模型自己造指令数据再学习' +description: '用 Self-Instruct 理解指令微调数据如何从少量种子任务扩展出来。' +来源: 'Wang et al., arXiv:2212.10560' +日期: 2026-07-14 +分类: LLM / Instruction Tuning +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2212.10560v2 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2212.10560 + source_version: arXiv:2212.10560v2 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v2 +--- + +## 是什么 + +Self-Instruct: Aligning Language Models with Self-Generated Instructions 是一篇 LLM / Instruction Tuning 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像老师先给 100 道样题，再让学生自己仿写 1 万道练习，老师负责筛掉重复和坏题。 + +它在本轮 40 篇里的位置是 **Batch 3 / instruction tuning**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +指令微调需要大量任务和答案，但人工写数据昂贵；只靠少量手工任务又覆盖不够广。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| Seed instructions | 从少量人工任务开始。 | +| 模型生成新任务 | 让 LLM 扩写 instruction、input 和 output。 | +| 过滤与微调 | 去重、筛质量，再训练模型跟随指令。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +给模型三个种子任务：翻译、摘要、分类。它生成“把会议纪要改写成待办清单”等新任务，再过滤相似样本，最后形成更大的 instruction 数据集。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **模型自举会放大原模型偏差**：模型自举会放大原模型偏差，生成数据不是天然干净。 +2. **过滤规则太弱会留下重复、空泛或错误答案。**：过滤规则太弱会留下重复、空泛或错误答案。 +3. **自生成数据提升 instruction following**：自生成数据提升 instruction following，但不保证事实性。 +4. **评测时要避开训练任务泄漏**：评测时要避开训练任务泄漏，否则会高估泛化。 + +## 学到什么 + +- 指令数据可以从“手工标注”扩展到“生成 + 过滤”的数据工程。 +- Self-Instruct 是 Alpaca、WizardLM 等数据路线的重要前身。 +- 数据生成流水线本身需要审计，而不只是看最终模型分数。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[wizardlm-2023]]、[[natural-instructions-v2-2022]]、[[instructgpt]]、[[flan-2021]] + +## 关联 + +- [[wizardlm-2023]] +- [[natural-instructions-v2-2022]] +- [[instructgpt]] +- [[flan-2021]] + +## 反向链接 + + diff --git a/src/content/docs/papers/speculative-decoding-2022.md b/src/content/docs/papers/speculative-decoding-2022.md new file mode 100644 index 000000000..d66f2e4a7 --- /dev/null +++ b/src/content/docs/papers/speculative-decoding-2022.md @@ -0,0 +1,90 @@ +--- +title: 'Speculative Decoding — 小模型先猜，大模型只验收' +description: '用 Speculative Decoding 理解如何不改变分布地加速自回归生成。' +来源: 'Leviathan et al., arXiv:2211.17192' +日期: 2026-07-14 +分类: LLM / Inference +难度: 高级 +difficulty: advanced +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2211.17192v2 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2211.17192 + source_version: arXiv:2211.17192v2 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v2 +--- + +## 是什么 + +Fast Inference from Transformers via Speculative Decoding 是一篇 LLM / Inference 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像助理先帮主编拟好接下来几句话，主编快速圈掉不合适的，保留合格部分。 + +它在本轮 40 篇里的位置是 **Batch 8 / long context and inference**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +LLM 自回归生成一次只确认一个 token，大模型推理延迟高，但很多位置小模型也能猜中。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| Draft model | 小模型一次提出多个候选 token。 | +| Target verification | 大模型并行验证这些候选是否可接受。 | +| 分布保持 | 通过接受/拒绝规则保证输出分布不被近似破坏。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +大模型要生成“今天的天气很好”。小模型先猜“的 天气 很”，大模型一次检查多个 token，猜对就跳过逐字生成。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **小模型太弱会猜不中**：小模型太弱会猜不中，反而增加开销。 +2. **任务越随机**：任务越随机，接受率越低。 +3. **实现需要高效批量验证和缓存管理。**：实现需要高效批量验证和缓存管理。 +4. **它加速 decoding**：它加速 decoding，不减少训练成本。 + +## 学到什么 + +- 推理优化可以利用“便宜模型预测，昂贵模型裁决”的结构。 +- Speculative Decoding 是服务端 LLM 延迟优化的基础技巧。 +- 后续 Medusa、EAGLE 等多 token 预测方法都在扩展这条路。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[medusa-2024]]、[[eagle]]、[[paged-attention]]、[[mistral-7b-2023]] + +## 关联 + +- [[medusa-2024]] +- [[eagle]] +- [[paged-attention]] +- [[mistral-7b-2023]] + +## 反向链接 + + diff --git a/src/content/docs/papers/star-self-taught-reasoner-2022.md b/src/content/docs/papers/star-self-taught-reasoner-2022.md new file mode 100644 index 000000000..328784a2e --- /dev/null +++ b/src/content/docs/papers/star-self-taught-reasoner-2022.md @@ -0,0 +1,90 @@ +--- +title: 'STaR — 模型先试着讲理由，再用对的理由训练自己' +description: '用 STaR 理解 rationale bootstrapping 怎样减少人工推理标注。' +来源: 'Zelikman et al., arXiv:2203.14465' +日期: 2026-07-14 +分类: LLM / Reasoning +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2203.14465v2 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2203.14465 + source_version: arXiv:2203.14465v2 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v2 +--- + +## 是什么 + +STaR: Bootstrapping Reasoning With Reasoning 是一篇 LLM / Reasoning 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像学生先自己写解题过程，老师只挑答案对且过程说得通的作业放进优秀范例本。 + +它在本轮 40 篇里的位置是 **Batch 4 / reasoning prompts**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +高质量 rationale 标注很贵，但没有中间理由，模型又难学会复杂推理。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| 生成 rationales | 让模型为训练题自己写推理过程。 | +| 答案过滤 | 只保留能得到正确答案的推理。 | +| 迭代微调 | 用筛过的 rationale 继续训练，再生成更好理由。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +给模型 100 道选择题，它先写理由和答案。只把答对的 40 道理由留下微调，下一轮可能答对更多。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **答对不代表理由真实**：答对不代表理由真实，可能是碰巧或事后合理化。 +2. **过滤会偏向模型已经会的题**：过滤会偏向模型已经会的题，困难题可能长期学不到。 +3. **迭代可能放大错误风格**：迭代可能放大错误风格，需要人工抽检。 +4. **rationale 数据可能泄漏答案线索**：rationale 数据可能泄漏答案线索，评测要谨慎。 + +## 学到什么 + +- 推理数据可以用“生成-过滤-再训练”自举出来。 +- STaR 是 Self-Instruct 在 reasoning 维度的近亲。 +- 任何自举流程都要警惕“看起来会解释”的幻觉。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[chain-of-thought]]、[[self-instruct-2022]]、[[gsm8k-2021]]、[[orca-explanation-tuning-2023]] + +## 关联 + +- [[chain-of-thought]] +- [[self-instruct-2022]] +- [[gsm8k-2021]] +- [[orca-explanation-tuning-2023]] + +## 反向链接 + + diff --git a/src/content/docs/papers/textual-inversion-2022.md b/src/content/docs/papers/textual-inversion-2022.md new file mode 100644 index 000000000..0ad5dadf2 --- /dev/null +++ b/src/content/docs/papers/textual-inversion-2022.md @@ -0,0 +1,90 @@ +--- +title: 'Textual Inversion — 给新概念学一个专属 token' +description: '用 Textual Inversion 理解冻结扩散模型时如何只学习概念 embedding。' +来源: 'Gal et al., arXiv:2208.01618' +日期: 2026-07-14 +分类: Diffusion / Personalization +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2208.01618v1 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2208.01618 + source_version: arXiv:2208.01618v1 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v1 +--- + +## 是什么 + +An Image is Worth One Word: Personalizing Text-to-Image Generation using Textual Inversion 是一篇 Diffusion / Personalization 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像给家里的猫起一个只有模型懂的外号，以后 prompt 里写这个外号就能召回它的视觉特征。 + +它在本轮 40 篇里的位置是 **Batch 9 / controllable generation**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +用户想把一个新物体、新风格或新人物概念加入模型，但不想全量微调扩散模型。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| 冻结生成模型 | 不更新 diffusion backbone。 | +| 学习新 token embedding | 用少量图片优化一个伪词向量。 | +| 组合式 prompt | 把新 token 和已有文本描述组合使用。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +给 3 张手工陶杯照片，学习 ``，之后写“a watercolor painting of on a desk”。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **一个 embedding 容量有限**：一个 embedding 容量有限，复杂主体可能学不完整。 +2. **训练图太少会过拟合**：训练图太少会过拟合，太杂会概念漂移。 +3. **组合能力取决于底座模型原有知识。**：组合能力取决于底座模型原有知识。 +4. **token 文件传播也可能携带未经授权的主体特征。**：token 文件传播也可能携带未经授权的主体特征。 + +## 学到什么 + +- Textual Inversion 是“只调控制向量，不改模型”的图像版 PEFT。 +- 它比 DreamBooth 更轻，但表达能力也更有限。 +- 个性化生成的最小交付物可以只是一个 embedding。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[dreambooth-2022]]、[[prompt-tuning-2021]]、[[controlnet-2023]]、[[edm-2022]] + +## 关联 + +- [[dreambooth-2022]] +- [[prompt-tuning-2021]] +- [[controlnet-2023]] +- [[edm-2022]] + +## 反向链接 + + diff --git a/src/content/docs/papers/toolllm-2023.md b/src/content/docs/papers/toolllm-2023.md new file mode 100644 index 000000000..97d96e58c --- /dev/null +++ b/src/content/docs/papers/toolllm-2023.md @@ -0,0 +1,90 @@ +--- +title: 'ToolLLM — 用 16000+ API 训练模型进入真实工具世界' +description: '用 ToolLLM 理解大规模 API 数据集、工具检索和工具评测如何支撑 agent。' +来源: 'Qin et al., arXiv:2307.16789' +日期: 2026-07-14 +分类: LLM / Tool Use +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2307.16789v2 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2307.16789 + source_version: arXiv:2307.16789v2 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v2 +--- + +## 是什么 + +ToolLLM: Facilitating Large Language Models to Master 16000+ Real-world APIs 是一篇 LLM / Tool Use 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像让新人客服接入一整个 SaaS 市场，不是背 10 个按钮，而是学会按需求找工具、读参数、处理返回。 + +它在本轮 40 篇里的位置是 **Batch 6 / agent tool ecosystems**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +工具调用研究常用少量手工 API，和真实世界成千上万接口的复杂性不匹配。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| ToolBench | 构造覆盖大量真实 API 的指令数据。 | +| API retriever | 先从工具池里找候选 API。 | +| ToolEval | 用自动和人工方式评估工具调用轨迹。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +用户要“查航班并订酒店”，系统要先找 flight search、hotel booking、calendar 等 API，再决定调用顺序，而不是只填一个函数。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **真实 API 会变更**：真实 API 会变更，静态数据集很快过期。 +2. **工具调用成功还要看鉴权、额度和错误处理。**：工具调用成功还要看鉴权、额度和错误处理。 +3. **评估轨迹比评估单次函数名更难。**：评估轨迹比评估单次函数名更难。 +4. **API 描述质量会影响 retriever 和 planner。**：API 描述质量会影响 retriever 和 planner。 + +## 学到什么 + +- ToolLLM 把 agent 评测从玩具函数推进到大工具池。 +- 工具规模上来后，检索、规划、执行和恢复必须分层。 +- 它和 Gorilla、MCP benchmark 是同一条工具可靠性主线。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[gorilla-2023]]、[[toolformer]]、[[mcpworld-2025]]、[[mcp-bench-2025]] + +## 关联 + +- [[gorilla-2023]] +- [[toolformer]] +- [[mcpworld-2025]] +- [[mcp-bench-2025]] + +## 反向链接 + + diff --git a/src/content/docs/papers/toxigen-2022.md b/src/content/docs/papers/toxigen-2022.md new file mode 100644 index 000000000..a7826177f --- /dev/null +++ b/src/content/docs/papers/toxigen-2022.md @@ -0,0 +1,90 @@ +--- +title: 'ToxiGen — 用生成模型造隐性仇恨测试集' +description: '用 ToxiGen 理解安全评测为什么要覆盖隐性、对抗性和群体相关文本。' +来源: 'Hartvigsen et al., arXiv:2203.09509' +日期: 2026-07-14 +分类: LLM / Safety Evaluation +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2203.09509v4 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2203.09509 + source_version: arXiv:2203.09509v4 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v4 +--- + +## 是什么 + +ToxiGen: A Large-Scale Machine-Generated Dataset for Adversarial and Implicit Hate Speech Detection 是一篇 LLM / Safety Evaluation 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像安全演练不只测明显脏话，还要测拐弯抹角、带暗示的攻击。 + +它在本轮 40 篇里的位置是 **Batch 10 / evaluation and safety**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +很多 toxic language 数据集偏向显性辱骂，模型可能漏掉更隐蔽、更接近真实平台风险的表达。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| 机器生成候选 | 用语言模型生成针对群体的隐性 toxic 文本。 | +| 人工筛选标注 | 对生成样本做质量和毒性判断。 | +| 对抗评测 | 检查分类器在隐性仇恨上的鲁棒性。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +“我不讨厌某群体，只是他们不适合某职业”这种句子没有明显脏词，却可能构成刻板印象攻击。ToxiGen 就关注这类样本。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **生成 toxic 文本本身要严格控制访问和用途。**：生成 toxic 文本本身要严格控制访问和用途。 +2. **群体标签和文化语境会影响标注一致性。**：群体标签和文化语境会影响标注一致性。 +3. **安全分类器可能误伤 reclaimed language 或讨论性文本。**：安全分类器可能误伤 reclaimed language 或讨论性文本。 +4. **数据集不能替代上线后的申诉和人工审核机制。**：数据集不能替代上线后的申诉和人工审核机制。 + +## 学到什么 + +- 安全评测要主动覆盖隐性风险，而不是只查关键词。 +- ToxiGen 展示了生成模型也能用于构造安全压力测试。 +- 越强的生成能力越需要配套治理和审计。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[truthfulqa-2021]]、[[constitutional-ai]]、[[promptfoo]]、[[toxigen-2022]] + +## 关联 + +- [[truthfulqa-2021]] +- [[constitutional-ai]] +- [[promptfoo]] +- [[toxigen-2022]] + +## 反向链接 + + diff --git a/src/content/docs/papers/truthfulqa-2021.md b/src/content/docs/papers/truthfulqa-2021.md new file mode 100644 index 000000000..970d70b01 --- /dev/null +++ b/src/content/docs/papers/truthfulqa-2021.md @@ -0,0 +1,90 @@ +--- +title: 'TruthfulQA — 专门问模型容易学人类谬误的问题' +description: '用 TruthfulQA 理解语言模型为什么会模仿常见假话而不是坚持事实。' +来源: 'Lin et al., arXiv:2109.07958' +日期: 2026-07-14 +分类: LLM / Evaluation +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2109.07958v2 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2109.07958 + source_version: arXiv:2109.07958v2 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v2 +--- + +## 是什么 + +TruthfulQA: Measuring How Models Mimic Human Falsehoods 是一篇 LLM / Evaluation 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像考试故意出“大家都误以为”的陷阱题，测学生是背流行说法还是查事实。 + +它在本轮 40 篇里的位置是 **Batch 10 / evaluation and safety**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +语言模型从互联网学习，可能把高频但错误的人类说法也学进去。流畅回答不等于真实回答。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| 对抗性问题集 | 收集容易诱发常见误解的问题。 | +| Truthfulness + informativeness | 同时看是否真实和是否有用。 | +| 模型规模分析 | 观察更大模型是否更容易模仿假话。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +问“如果你吞下口香糖，它会在胃里停留七年吗？”模型若复述都市传说就失败，必须纠正常见误解。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **TruthfulQA 覆盖的是特定错误类型**：TruthfulQA 覆盖的是特定错误类型，不代表全面事实性。 +2. **回答保守可能更 truthful 但不够 informative。**：回答保守可能更 truthful 但不够 informative。 +3. **评测题会逐渐进入训练数据**：评测题会逐渐进入训练数据，需版本化管理。 +4. **真实产品还要处理来源引用和时效性。**：真实产品还要处理来源引用和时效性。 + +## 学到什么 + +- 事实性不是语言流畅度的副产品。 +- TruthfulQA 提醒我们评测要主动找模型会犯的“人类式错误”。 +- RAG、引用和拒答策略都可以看作对这类问题的工程回应。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[webgpt-2021]]、[[rag-lewis-2020]]、[[constitutional-ai]]、[[toxigen-2022]] + +## 关联 + +- [[webgpt-2021]] +- [[rag-lewis-2020]] +- [[constitutional-ai]] +- [[toxigen-2022]] + +## 反向链接 + + diff --git a/src/content/docs/papers/ul2-2022.md b/src/content/docs/papers/ul2-2022.md new file mode 100644 index 000000000..9ac28169e --- /dev/null +++ b/src/content/docs/papers/ul2-2022.md @@ -0,0 +1,90 @@ +--- +title: 'UL2 — 一个模型同时练完补空、续写和长文本' +description: '用 UL2 理解 mixture-of-denoisers 如何统一不同语言模型训练范式。' +来源: 'Tay et al., arXiv:2205.05131' +日期: 2026-07-14 +分类: LLM / Pretraining Objective +难度: 高级 +difficulty: advanced +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2205.05131v3 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2205.05131 + source_version: arXiv:2205.05131v3 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v3 +--- + +## 是什么 + +UL2: Unifying Language Learning Paradigms 是一篇 LLM / Pretraining Objective 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像一名学生同时练填空题、作文续写和长篇阅读，不再被单一题型绑住。 + +它在本轮 40 篇里的位置是 **Batch 1 / foundation scaling**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +BERT 式 denoising、T5 式 span corruption、GPT 式 causal LM 各有优势。问题是能不能用一套目标让模型学到多种模式，而不是为每种任务训练一支模型。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| Mixture-of-Denoisers | 把 R-denoising、S-denoising、X-denoising 混在预训练里。 | +| Mode token | 让模型知道当前该按哪种破坏/恢复模式工作。 | +| 统一 encoder-decoder 视角 | 在理解、生成和长上下文任务之间做折中。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +同一段话，训练时有时遮住少量词，有时遮住长 span，有时只给前半段让模型续写。模型被迫学会“补局部”和“接全局”两种能力。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **混合目标不是越多越好**：混合目标不是越多越好，比例不当会互相拉扯。 +2. **mode token 提供了控制面**：mode token 提供了控制面，但也让部署 prompt 多一层心智负担。 +3. **UL2 的结论和具体架构、数据、规模绑定**：UL2 的结论和具体架构、数据、规模绑定，不能直接套到任意模型。 +4. **统一训练范式不等于统一产品接口**：统一训练范式不等于统一产品接口，推理时仍要选择合适模式。 + +## 学到什么 + +- 预训练目标本身就是产品能力的上游开关。 +- UL2 把“理解 vs 生成”的二分变成可组合设计。 +- 后续 Flan、T5 系和指令模型都可以从这个角度理解。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[t5]]、[[bert]]、[[palm-2022]]、[[prompt-tuning-2021]] + +## 关联 + +- [[t5]] +- [[bert]] +- [[palm-2022]] +- [[prompt-tuning-2021]] + +## 反向链接 + + diff --git a/src/content/docs/papers/webgpt-2021.md b/src/content/docs/papers/webgpt-2021.md new file mode 100644 index 000000000..9f23704e0 --- /dev/null +++ b/src/content/docs/papers/webgpt-2021.md @@ -0,0 +1,90 @@ +--- +title: 'WebGPT — 让模型带着浏览器回答问题' +description: '用 WebGPT 理解检索、引用和人类偏好如何组合成可追溯问答。' +来源: 'Nakano et al., arXiv:2112.09332' +日期: 2026-07-14 +分类: LLM / Browser Agent +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2112.09332v3 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2112.09332 + source_version: arXiv:2112.09332v3 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v3 +--- + +## 是什么 + +WebGPT: Browser-assisted question-answering with human feedback 是一篇 LLM / Browser Agent 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像开卷考试：学生可以查网页，但必须把引用贴出来，还要让老师判断答案是否真正支持结论。 + +它在本轮 40 篇里的位置是 **Batch 5 / agents and tools**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +纯参数问答容易编造事实；检索系统能找资料，但不一定会组织成自然答案并标注依据。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| 浏览器动作空间 | 模型可以搜索、打开页面、引用片段。 | +| 示范与偏好学习 | 先学人类浏览轨迹，再用偏好优化答案。 | +| 带引用回答 | 输出答案时附上可检查来源。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +问“某论文是哪年发表的”，WebGPT 式 agent 会搜索标题、打开可信页面、引用出版信息，而不是凭记忆猜年份。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **引用存在不代表支持结论**：引用存在不代表支持结论，仍要检查 claim-source 对齐。 +2. **搜索结果会受排名和网页质量影响。**：搜索结果会受排名和网页质量影响。 +3. **浏览轨迹成本高**：浏览轨迹成本高，实时产品要控制步数。 +4. **人类偏好可能偏向流畅答案**：人类偏好可能偏向流畅答案，而不是最严谨答案。 + +## 学到什么 + +- WebGPT 是 RAG、browser agent 和 citation QA 的早期汇合点。 +- 可追溯回答需要动作记录、来源和偏好训练一起工作。 +- 今天的 AI 搜索产品仍在解决同一个 claim grounding 问题。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[rag-lewis-2020]]、[[graphrag]]、[[truthfulqa-2021]]、[[react-agent]] + +## 关联 + +- [[rag-lewis-2020]] +- [[graphrag]] +- [[truthfulqa-2021]] +- [[react-agent]] + +## 反向链接 + + diff --git a/src/content/docs/papers/wizardlm-2023.md b/src/content/docs/papers/wizardlm-2023.md new file mode 100644 index 000000000..e07ccf334 --- /dev/null +++ b/src/content/docs/papers/wizardlm-2023.md @@ -0,0 +1,90 @@ +--- +title: 'WizardLM — 用 Evol-Instruct 自动变难训练题' +description: '用 WizardLM 理解 instruction 数据不只要多，还要逐步变复杂。' +来源: 'Xu et al., arXiv:2304.12244' +日期: 2026-07-14 +分类: LLM / Instruction Tuning +难度: 中级 +difficulty: intermediate +trust: + version: study-v2 + source_kind: paper + note_type: paper + canonical_source: https://arxiv.org/abs/2304.12244v3 + source_authority: AUTHOR_PRIMARY + accessed_at: '2026-07-14' + publication_id: arXiv:2304.12244 + source_version: arXiv:2304.12244v3 + evidence_type: STATIC_ANALYSIS + verification_status: UNVERIFIED + reviewed_at: '2026-07-14' + review_after: null + applicable_version: arXiv v3 +--- + +## 是什么 + +WizardLM: Empowering large pre-trained language models to follow complex instructions 是一篇 LLM / Instruction Tuning 论文。本卡只基于 arXiv 官方元数据和论文静态阅读做研究整理；没有运行作者代码，也没有复现论文分数。 + +类比：像刷题系统会把“写一句话总结”升级成“按三种受众写三版摘要并比较差异”，难度被系统性拉高。 + +它在本轮 40 篇里的位置是 **Batch 3 / instruction tuning**：不是孤立收藏，而是补上 study 论文图谱里还缺的一块。 + +## 问题是什么 + +很多 instruction 数据停留在简单任务，模型会变得听话但不擅长复杂约束。 + +如果把它放进产品工程语境，核心问题是：团队到底应该把不确定性留给模型本身，还是拆给数据、工具、训练目标、评测和系统约束分别处理。 + +## 为什么重要 + +- 它给后续研究提供了一个可引用的名字和问题边界。 +- 它把一个模糊能力拆成了可以讨论的机制或流程。 +- 它提醒我们不要只看最终 benchmark，而要看数据、约束和验收方式。 +- 它能和本库已有笔记形成交叉链接，方便以后按主题复习。 + +## 核心方法 + +| 设计 | 作用 | +|---|---| +| Evol-Instruct | 用 LLM 自动改写指令，让深度、广度和约束增加。 | +| 复杂任务微调 | 把演化后的任务用于 instruction tuning。 | +| 人工/模型评估 | 比较复杂指令上的遵循能力。 | + +这三点合在一起，给这篇论文建立了一个最小可理解模型：先看它把问题切在哪里，再看它把哪部分交给模型、哪部分交给外部结构。 + +## 手工 toy 复现 + +原任务是“解释二分查找”；演化后变成“给零基础同学解释二分查找，列两个误区，再给一道练习题”。模型由此学会处理多约束输出。 + +这个 toy 复现只验证机制直觉，不声明论文原始指标已复现。真正升级为 VERIFIED 需要独立执行证据和 review receipt 绑定。 + +## 踩过的坑 + +1. **自动变难可能制造不自然任务**：自动变难可能制造不自然任务，和真实用户需求脱节。 +2. **复杂指令越长**：复杂指令越长，答案质量越难自动判定。 +3. **演化数据依赖基础模型能力**：演化数据依赖基础模型能力，弱模型会产生坏任务。 +4. **只优化复杂性可能牺牲简洁回答能力。**：只优化复杂性可能牺牲简洁回答能力。 + +## 学到什么 + +- 指令数据的“难度曲线”本身是训练设计对象。 +- WizardLM 是从 Self-Instruct 到复杂 agent 任务数据的桥。 +- 产品评测也应分层：简单 obey、复杂约束、跨步执行要分开看。 + +## 延伸阅读 + +- 原文： +- 本卡使用版本： +- 主题关联：[[self-instruct-2022]]、[[natural-instructions-v2-2022]]、[[orca-explanation-tuning-2023]]、[[toolllm-2023]] + +## 关联 + +- [[self-instruct-2022]] +- [[natural-instructions-v2-2022]] +- [[orca-explanation-tuning-2023]] +- [[toolllm-2023]] + +## 反向链接 + + diff --git a/src/content/docs/queue.md b/src/content/docs/queue.md index 886db47cd..21570f0ba 100644 --- a/src/content/docs/queue.md +++ b/src/content/docs/queue.md @@ -6,7 +6,7 @@ sidebar: --- > 不是"读哪 20 个"的清单，是"先读哪 5 个就能撑起一个领域"的导航。 -> 当前站点 961 篇项目笔记 + 1023 篇论文笔记，凑数没有意义，**取舍**才有。 +> 当前站点 961 篇项目笔记 + 1063 篇论文笔记，凑数没有意义，**取舍**才有。 > 每个主题给 3-5 个 pillar：反向链接最多、跨主题被引最广、读完能形成判断。 ## 怎么用这页 @@ -189,7 +189,7 @@ PL 理论在论文侧根扎得最深：[[hindley-milner]] / [[lambda-calculus]] ## 全景 atlas - 项目全景（961 篇按主题分组、反向链接热度、消化状态）：[projects-atlas](/study/projects-atlas/) -- 论文全景（1023 篇按子领域、pillar 标记、未消化队列）：[papers-atlas](/study/papers-atlas/) +- 论文全景（1063 篇按子领域、pillar 标记、未消化队列）：[papers-atlas](/study/papers-atlas/) - 论文推荐入口（与本页平行的论文版导航）：[papers-queue](/study/papers-queue/) - 方法论与挑选标准：[about](/study/about/) / [method](/study/method/) / [papers-method](/study/papers-method/)