S1CogBench
by Dengzhe Hou and Kazunori D. Yamada
What is S1CogBench? S1CogBench is a reasoning & knowledge eval made by Dengzhe Hou and Kazunori D. Yamada. Puts fast "System One" decision models through 24 established experiments from cognitive science, run in 27 evaluation modes. Each model is checked three ways: how well it predicts the choices people actually made, how it does when it plays the task itself, and how its stated probabilities behave, with controlled manipulations of the tasks to see whether it reacts as people do. The paper evaluates seven open models plus one API model and finds that good scores on some learning tasks sit alongside failures to gather information or adjust to feedback, and that rankings partly depend on how choices are read out. The repository ships the original runners and scorers, a quick smoke test and a guide for adding a model; some modes need the gated Psych-101 test set (arXiv:2611.00911).
Editor note
Puts fast "System One" decision models through 24 established experiments from cognitive science, run in 27 evaluation modes. Each model is checked three ways: how well it predicts the choices people actually made, how it does when it plays the task itself, and how its stated probabilities behave, with controlled manipulations of the tasks to see whether it reacts as people do. The paper evaluates seven open models plus one API model and finds that good scores on some learning tasks sit alongside failures to gather information or adjust to feedback, and that rankings partly depend on how choices are read out. The repository ships the original runners and scorers, a quick smoke test and a guide for adding a model; some modes need the gated Psych-101 test set (arXiv:2611.00911).
- What it tests
- Cognitive-experiment behaviour of fast decision models: human-choice prediction, autonomous play and explicit probability reports.
- Original source
- github.com/dengzhe-hou/S1CogBench
- Last updated
- arXiv:2611.00911 (14 Nov 2026); repo last pushed 2026-11-20 (GitHub API, read 22 Nov 2026)
- License
- MIT (code, GitHub API, read 22 Nov 2026); aggregate results CC-BY-4.0 (HF tag)
Related reasoning & knowledge evals
- GPQA — Expert-level science reasoning.
- Humanity's Last Exam — Frontier academic knowledge and reasoning.
- LiveBench — Hard, contamination-limited tasks with verifiable ground-truth answers (no LLM judge) in v1; LiveBench 2.0 (14 Oct 2026) adds long-horizon real-work tasks.
- MMLU — World knowledge and problem solving across many school and professional subjects.
Discussion
No one has discussed S1CogBench yet. Have you run it? Share what it caught, what it missed or how you set it up.