AI Paper Atlas

Research map / Large language models / Agents, code and math

NewtonBench: Benchmarking Generalizable Scientific Law Discovery in LLM Agents

Tianshi Zheng, Kelvin Kiu Wai Tam, Newt Nguyen Kim Hue Nam, Baixuan Xu, Zhaowei Wang, Cheng Jiayang, Hong Ting Tsang, Weiqi Wang, Jiaxin Bai, Tianqing Fang, Yangqiu Song, Ginny Wong and 1 more

ICLR 2026 · Large language models › Agents, code and math · approach: Agent instruction-following benchmarks · 32 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
give the weaker agent a code interpreterGPT-4.1-mini · NewtonBench avgsymbolic accuracy (%): 5.1 -> 14.7Improved
give the weaker agent a code interpreterGPT-4.1 · NewtonBench avgsymbolic accuracy (%): 5.8 -> 18.9Improved
give the weaker agent a code interpreterDeepSeek-V3 · NewtonBench avgsymbolic accuracy (%): 8.9 -> 13.3Improved
give the stronger agent a code interpreterGPT-5 · NewtonBench avgsymbolic accuracy (%): 75.9 -> 73.3Hurt
give the stronger agent a code interpreterGemini-2.5-flash · NewtonBench avgsymbolic accuracy (%): 50.6 -> 40.9Hurt
give the stronger agent a code interpreterGPT-5-mini · NewtonBench avgsymbolic accuracy (%): 53.1 -> 48.1Hurt
give the stronger agent a code interpreterGemini-2.5-pro · NewtonBench avgsymbolic accuracy (%): 65.4 -> 63.7Hurt
give the agent a code interpreterQwQ-32B · NewtonBench avgsymbolic accuracy (%): 28.5 -> 31.4Improved
use a reasoning model instead of non-reasoning (GPT-4.1-mini to GPT-5)GPT family · NewtonBench avgsymbolic accuracy (%): 5.1 -> 75.9Improved
increase system complexity from vanilla equation to complex system, hard levelGPT-5 · NewtonBenchsymbolic accuracy (%): 87.5 -> 40.3Hurt
add Gaussian observation noise of 0.0001GPT-5-mini · NewtonBenchsymbolic accuracy: 12-16% reduction in accuracyHurt
increase noise level from 0.0001 to 0.1GPT-5-mini · NewtonBenchRMSLE: RMSLE remained relatively stableNo clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy