Model, Not Mode: a pre-registered comparison of fast and reasoning modes across six LLM families, using 60 chat-app answers to one constrained systems-design task, blind scoring and variance decomposition. Includes data, analysis code and reproducible results.
python benchmark reproducible-research reasoning preregistration variance-decomposition systems-architecture large-language-models llm-evaluation llm-as-a-judge test-time-compute reasoning-mode
-
Updated
Oct 28, 2026 - TeX