Skip to content
View Yushu-Li's full-sized avatar
16 followers · 25 following

Achievements

Achievement: Pull SharkAchievement: Starstruck

Achievements

Achievement: Pull SharkAchievement: Starstruck

Highlights

  • Pro

Block or report Yushu-Li

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Yushu-Li/README.md

Yushu Li — Research. Build. Iterate. Learning systems that adapt, reason, and improve.

Applied Scientist Intern @ AWS
PhD Student @ UBC ECE  ·  Student Researcher @ Vector Institute

Website & Publications ↗    /    Google Scholar ↗    /    LinkedIn ↗    /    CV ↗

I’m Yushu, an AI researcher interested in how learning systems adapt, reason, and improve through feedback. At AWS, I work on robust and efficient reinforcement learning. I also enjoy turning ideas into useful tools and contributing fixes to software people use.

01 / Research interests

◈   Reward Hacking

Understanding when optimizing a reward misses the intended goal—and how to make reinforcement learning more reliable.

REWARD DESIGN   /   RELIABLE RL

↗   Efficient AI

Getting more useful reasoning and learning from limited computation, with a focus on budget-aware agents and efficient RL.

REASONING BUDGETS   /   EFFICIENCY

⌘   LLM Agents & Reasoning

Exploring tool use, search, and evaluation to help agents solve real tasks and learn from their outcomes.

TOOL USE   /   SEARCH   /   FEEDBACK

◎   Adaptation & Robustness

Helping models handle distribution shifts and unfamiliar classes through test-time adaptation and vision-language learning.

TEST-TIME ADAPTATION   /   OPEN WORLD

02 / Selected work

ECALP — ICLR 2025. Context-aware label propagation for vision-language models.

Adapt without retraining.
Context-aware label propagation for zero- and few-shot adaptation of vision-language models.

Explore the code ↗ 		OWTTT — ICCV 2023 Oral. Open-world test-time training.

Adapt to an open world.
Robust test-time training with self-training and dynamic prototype expansion.

Explore the code ↗

More research on reward hacking, agentic RL, and efficient reasoning lives on my publications page ↗.

03 / Currently exploring

  • Agent RSI — recursive self-improvement through feedback, evaluation, and iteration.
  • Jev — structured, probabilistic decisions and how they fit into agent systems.
  • AI as an enabler — small tools and automations that make something useful possible with less effort.

04 / Open source, in practice

I like contributions with a concrete outcome: a bug fixed, a workflow simplified, or an experiment made easier to run.

OpenEvolve · Selected merged contributions

Contribution What it enables
Evaluator interoperability · #519 Running local AlphaEvolve evaluators with OpenEvolve.
Experiment configuration import · #520 Reusing AlphaEvolve experiment configurations in OpenEvolve.

More merged contributions ↗

Outside research, I enjoy experimenting with AI-assisted video creation and automating repetitive work—using AI to gain new capabilities and leave something useful behind.

Build something small. Make it work. Get a real result. Then iterate.

Open to conversations about research, practical AI systems, and useful open source.

Pinned Loading

  1. OWTTT OWTTT Public

    [ICCV 2023 Oral] Official repository for “On the Robustness of Open-World Test-Time Training: Self-Training with Dynamic Prototype Expansion”

    Python 46 4

  2. ECALP ECALP Public

    [ICLR 2025] Official repository for “Efficient and Context-Aware Label Propagation for Zero-/Few-Shot Training-Free Adaptation of Vision-Language Model”

    Python 15