Can LLMs Uncover Hidden Physical Laws through Free Exploration?

Zhen Yang; Wen Wang; Xander Xu; Ying-Cong Chen

cs.CLOct 27, 2026

Can LLMs Uncover Hidden Physical Laws through Free Exploration?

Authors: Zhen Yang, Wen Wang, Xander Xu, Ying-Cong Chen

Organizations: HKUST(GZ) · ZJU · Alibaba Group · HKUST

Abstract

Scientific discovery often begins with an unexpected observation that prompts a new research question. In a world with altered physical laws, can large language models (LLMs) notice unexpected behavior, such as heavier objects falling more slowly, and discover the law through experiments without being told what to investigate? To evaluate this ability, we introduce LongChang, an automated framework that creates simulated worlds with altered physics, where one or more laws differ from ordinary physics. The evaluated LLMs act as explorers, manipulating objects and building tools to investigate their own questions without research tasks or hints about these laws. Separate LLM agents generate and review scenes, review tool designs, and evaluate reports for correctness and experimental support using the altered physical laws and exploration records. These records guide scene evolution for further exploration, allowing the full cycle to run without human intervention. Among 12 LLMs, GPT-6-Astra achieves a discovery rate of 16.7%, while the other 11 score 0%. The exploration records reveal uninvestigated anomalies and only partial identification of the altered physical laws.

Figures & tables

Panel of Table 1 — Table 1: Scene 01 uses ordinary physics. Scenes 02–15 each have one altered physical law. Laws are in App. A.1 .

Figure 1: LongChang’s glass room. Explorers manipulate objects and build tools. Other LLMs evolve scenes using exploration records without human intervention.

Closest work	No task	Interactive 3D scenes	Tool design	Scene evolution
PhysGym	✗	✗	✗	✗
NewtonBench	✗	✗	✗	✗
DiscoverPhysics	✗	✗	✗	✗
LongChang	✓	✓	✓	✓

Table 2: LongChang compared with the closest work: PhysGym ( Chen et al., 2025 ) , NewtonBench ( Zheng et al., 2026 ) , and DiscoverPhysics ( Wiemann et al., 2026 ) . No task: no assigned research goal. Interactive 3D scenes: physical worlds in MuJoCo ( Todorov et al., 2012 ) . Tool design: devices built to interact with scene objects. Scene evolution: automatically generating and revising scenes guided by exploration records.

Figure 2: LongChang’s automated cycle of scene construction, free exploration, assessment, and scene evolution.

Model	D	C	S	R	L	ES (%)	CS	CF (%)	AF (%)
	Discovery (%)			Control	Exploration
GPT-5.6-Luna	0.0	0.0	38.1	0	36.7	95.6	0	–	0.8
GPT-6-Astra	16.7	19.0	97.6	0	39.3	100.0	11	48.0	2.1
Claude-Opus-5	0.0	16.7	0.0	0	85.8	100.0	44	66.6	18.2
Gemini-3.7-Flash	0.0	0.0	19.0	0	10.9	100.0	1	75.0	6.1
GLM-5.3	0.0	4.8	0.0	3	91.6	97.8	45	64.0	16.8

Table 3: Discovery performance and exploration behavior of 12 LLMs in LongChang, with metrics from Sec. 3.4 .

Figure 5: Case study of research question formation in Scene 15. (a) GPT-6-Astra investigates red’s delayed restart and discovers the altered physical law. (b) Gemini-3.7-Flash observes the restart but still follows its original plan.

Figure 6: Case study of testing competing explanations.

Figure 7: Case study of tool construction and redesign.

Task guidance, 100 turns
Condition	D (%)	C (%)	S (%)	L	AF (%)
Free exploration	16.7	19.0	97.6	39.3	2.1
Task-guided	28.6	31.0	97.6	30.9	2.7
Turn budget
100 turns	16.7	16.7	100.0	42.2	2.7
500 turns	14.3	14.3	100.0	41.4	3.6

Table 4: Effects of task guidance, turn budget, and apparatus access on discovery and exploration behavior.

Panel of Table 5 — Table 5: Scene evolution guided by a corrected error in Scene 15, with altered physical laws in three new scenes.

Scene	C (%)	S (%)	D (%)
Scene 15	33.3	100.0	33.3
Scene 14	66.7	100.0	66.7
Generation	0.0	100.0	0.0
Refinement	0.0	100.0	0.0
Composition	0.0	100.0	0.0

Table 6: Discovery rates in source and evolved scenes.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Scene	Physical laws
01 Normal control	Follows ordinary physics (App. A.1 ).

02 Red attraction	Free single-sphere red bodies attract each other with $\mathbf{F}=20\mathbf{d}/\max(r,0.5)^{3}$ , where $\mathbf{d}$ points toward the other sphere and $r=\|\mathbf{d}\|$ .

03 Mass-dependent fall	Each free body has downward acceleration $g_{0}\sqrt{m_{0}/m}$ , with $m_{0}=1$ .

Appendix

Table 7: Altered physical laws of the scenes in Tab. 1 .

Scene	Physical laws
Shared conditions	Conventions follow Tab. 7 . Extra accelerations (in bricks/s 2 ) start active and add to ordinary physics. Each launch of any movable object restarts a 10-second pause.
Generation	When active, red and blue repel horizontally with acceleration 1.5, independent of separation. Repulsion is zero when their horizontal centers coincide.

Refinement	After the pause, red accelerates northward at 1.5 exactly when blue is outside the closed horizontal square $[2.5,3.5]\times[2.5,3.5]$ , regardless of blue’s height or held state. Blue starts outside. Crossing the boundary does not reset the timer.

Composition	Red follows Scene 15’s northward-force rule, while its launches also trigger Scene 14’s delayed velocity copy to green. The 10-second force pause and 12-second copy delay operate independently.

Appendix

Table 8: Altered physical laws of the evolved scenes in Sec. 5 .


Action	Target	Parameters	Effect
PUSH	Released movable object	Object ID, heading, impulse $[0,10]$	Adds a horizontal impulse to existing motion.
THROW	Released movable object	Object ID, heading, elevation $[0^{\circ},90^{\circ}]$ , impulse $[0,10]$	Adds a directed impulse to existing motion.

WAIT	Entire scene	Integer duration: 1–30 counts	Advances simulation time.

Appendix

Table 9: Targets, parameters, and effects of exploration actions.

AI Agents for Scientific Discovery

AI agents for scientific discovery use language models, tools, and experiments to generate, test, and refine scientific knowledge.

See all 288 papers

DiscoverPhysics: Benchmarking LLMs for Out-of-the-Box Scientific Thinking
May 25, 2026Matt L. Wiemann, Lindsay M. Smith, Peter Melchior +4Physical Reasoning Scientific Discovery
Princeton University · Work done during internship at NYU/Polymathic AI. · Boston University +3
Can LLMs Discover Scientific Laws in Real and Parallel Worlds?
Sep 1, 2026Yiming Huang, Ziche Liu, Junxia Cui +11Scientific Discovery Scientific Hypothesis Generation
University of California San Diego · University of California, Irvine · Cushing Academy +1
Agentic Exploration of Physics Models
Sep 29, 2025Maximilian Nägele, Florian MarquardtAI Agents for Scientific Discovery Scientific Discovery
Max Planck Institute for the Science of Light, Staudtstraße 2, 91058 Erlangen, Germany · Department of Physics, Friedrich-Alexander Universität Erlangen-Nürnberg, Staudtstraße 5, 91058 Erlangen, Germany

May 25, 2026J/K move · Enter open · S save

DiscoverPhysics: Benchmarking LLMs for Out-of-the-Box Scientific Thinking

Matt L. Wiemann, Lindsay M. Smith, Peter Melchior, Siddharth Mishra-Sharma+3

Princeton University · Work done during internship at NYU/Polymathic AI. · Boston University+3

Physical Reasoning Scientific Discovery AI Agent Benchmarks System Identification Large Language Model-Guided Scientific Discovery

Open paper AskPDF arXiv

Frontier LLMs now perform strongly across a wide range of physics evaluations, but it is hard to disentangle genuine reasoning from recall of established science. We introduce DiscoverPhysics, an interactive benchmark that asks a LLM agent to discover the laws of motion of a simulated world whose physics deliberately deviates from our own. We construct 22 worlds governed by, among others, screened and fractional-power gravity, multi-species couplings, hidden dark-matter-like particles, non-coordinate-free physics, and time-varying interactions. Each world is generated on demand by an N-body simulator, for which the agent proposes several rounds of experiments, observes raw trajectory data, and ultimately submits both a natural-language explanation of the world's physics and a Python implementation of the inferred law. Because solving a world requires the agent to design informative experiments and revise its hypotheses, the benchmark probes long-horizon reasoning over an experimental history. We evaluate submissions along two complementary axes: trajectory MSE on held-out particles and an LLM-judged explanation score following an expert-written rubric assessing conceptual understanding of each world. Across eleven frontier models, we find that the strongest agents pass only half of the worlds and consistently fail on those where latent structure must be uncovered. Open-source models lag substantially behind commercial models, both in their ability to design informative experiments and in extracting conclusions from the data. We further find that good predictive accuracy does not guarantee high explanation quality and that conceptual understanding depends on hypothesis refinement through well-chosen experiments.

Abstract

AI Agents for Scientific Discovery

Explore similar work

DiscoverPhysics: Benchmarking LLMs for Out-of-the-Box Scientific Thinking

Can LLMs Discover Scientific Laws in Real and Parallel Worlds?

Agentic Exploration of Physics Models