NewtonBench
About this tool
HKUST KnowComp's benchmark for scientific-law discovery by model agents. It systematically modifies classical laws, including gravitation, Coulomb's law and Fourier's law, to reduce memorized answers. Agents collect data through interactive experiments and infer the underlying laws. Its 324 tasks span 12 physics domains; the paper appeared at ICLR 2026.
- Feature 1: 324 law-discovery tasks across 12 physics domains.
- Feature 2: Modified classical laws reduce reliance on memorization.
- Feature 3: Difficulty levels cover law difficulty and system complexity.
- Feature 4: Evaluation settings for reasoning-only and code-assisted agents.
- Feature 5: Integrated into NVIDIA NeMo Gym as a reinforcement-learning environment.
Resource information
- Maintainer
- HKUST-KnowComp, Associate Professor Yangqiu Song's team; official university-laboratory project.
- Type
- Dataset / model
- Category
- Research agents & deep research
- Runs on
- Dataset or model
- Who it's for
- Researchers evaluating or training scientific-discovery agents, AI scientists and model reasoning, including comparisons of law-induction performance.
- License at time of listing
- MIT
- Source
- GitHub
