Can LLMs Uncover Hidden Physical Laws through Free Exploration?

Abstract: Scientific discovery often begins with an unexpected observation that prompts a new research question. In a world with altered physical laws, can 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.

• The paper demonstrates that among 12 evaluated LLMs, only GPT-6-Astra achieves any fully successful discoveries with a discovery rate of 16.7%, revealing the challenge and potential of autonomous scientific inquiry.
• GPT-6-Astra's success in automated experimentalism was isolated thanks to assessing both the exploration process and final conclusions, showing a strong ability to recognize patterns and formulate hypotheses, though the analysis emphasizes 'scientific inference' and documentation of evidence.
• Improvements in discovery rates can be attributed to the adaptive allocation of experimental efforts and the capacity to recognize when a planned experiment should be altered based on new observations.

Research question and contribution

“Can LLMs Uncover Hidden Physical Laws through Free Exploration?” (2610.37649) studies a stringent form of autonomous scientific inquiry: whether an LLM can notice an anomalous physical event, formulate an investigation without being assigned a research objective, design experiments that discriminate among explanations, and report an experimentally supported physical law.

The central distinction is between law discovery as task completion and law discovery as self-directed inquiry. Existing physical-law benchmarks generally specify the quantity or equation to recover. LongChang removes that specification. The explorer receives only an instruction to explore freely and report what it learns. It is not told that physics may be altered, which objects are relevant, or which anomaly should be investigated. The benchmark therefore evaluates the transition from observation to research question, rather than only equation fitting after a research target has been supplied.

The paper’s principal empirical result is stark: among 12 evaluated LLMs, only GPT-6-Astra achieves any fully successful discoveries, with a discovery rate of 16.7%. The other 11 models score 0%. The result is not attributable simply to a lack of interaction or tool access. Models often explore extensively, construct apparatus, or produce partially correct explanations, but fail to establish the complete law together with its scope, trigger conditions, and experimental basis.

LongChang’s environment and evaluation protocol

LongChang generates simulated physical worlds in MuJoCo. Its base suite contains one ordinary-physics control scene and 14 scenes, each containing one altered physical law. The laws modify diverse dependencies: object mass, spatial proximity, velocity, contact history, tile identity, delayed effects, object identity, and prior actions. Examples include mass-dependent falling acceleration, a speed limit, velocity scaling when entering and exiting a region, collision-induced velocity amplification, contact-triggered color exchange, delayed velocity copying, and a northward force that temporarily pauses after a launch.

The explorer interacts with objects through pushing, throwing, grasping, moving, releasing, waiting, and marking reference points. It may also construct physical apparatus by submitting MuJoCo XML specifications. An apparatus can contain bodies, joints, springs, tendons, and measurement sites, allowing the explorer to build instruments rather than relying exclusively on primitive manipulation.

The evaluation separates two criteria:

• Law correctness (C): whether the report identifies the altered law, its relevant objects, conditions, triggers, and identifiable parameters.
• Evidence support (S): whether the exploration record contains experiments sufficient to justify the report at the stated level of certainty.

A discovery requires both criteria. This separation is essential. A model can guess a correct law without demonstrating it, or document a genuine anomaly without resolving the underlying law.

LongChang also introduces an automated scene-evolution loop. LLM-based designers generate, refine, or compose scenes using unresolved questions and corrected errors extracted from prior exploration records. Separate agents validate scene programs, review apparatus, and assess reports. The resulting framework is therefore not merely a fixed benchmark; it is a closed pipeline connecting environment construction, exploration, evaluation, and subsequent challenge generation.

Figure 1: LongChang’s glass-room environment, where explorers manipulate objects and construct physical instruments while separate agents evolve scenes from exploration records.

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

Main empirical findings

GPT-6-Astra obtains a discovery rate of 16.7%, a law-correctness rate of 19.0%, and an evidence-support rate of 97.6%. This gap is the paper’s most informative quantitative finding. The model generally produces reports grounded in its records, but those reports usually identify only a fragment of the altered law. For example, it may establish that two red spheres attract without determining the distance dependence or whether the effect generalizes to other red objects. Thus, high evidence support does not imply complete scientific identification.

Claude-Opus-5 exhibits the opposite pattern: it reaches 16.7% law correctness but 0% evidence support. This indicates that correct-looking conclusions can arise without adequate experimental justification. The contrast between GPT-6-Astra and Claude-Opus-5 demonstrates why evaluating only final-answer correctness would conflate hypothesis generation with scientific validation.

Model	Discovery	Correctness	Support	Mean turns	Construction success
GPT-6-Astra	16.7%	19.0%	97.6%	39.3	11
GPT-5.6-Luna	0%	0%	38.1%	36.7	0
Claude-Opus-5	0%	16.7%	0%	85.8	44
Gemini-3.7-Flash	0%	0%	19.0%	10.9	1
GLM-5.3	0%	4.8%	0%	91.6	45
Kimi-K3	0%	2.4%	0%	83.9	42
Grok-4.7	0%	0%	2.4%	85.2	2

The behavioral statistics reject several simple explanations for failure. Gemini-3.7-Flash stops after an average of only 10.9 turns, which limits opportunities for follow-up experimentation. However, the GLM models use approximately 90 turns, construct apparatus successfully in every run, and still achieve no discoveries. Their construction-failure rates exceed 60%, and failed actions account for 16.8–20.2% of turns. Grok-4.7 is comparatively persistent but also makes no discoveries. Consequently, persistence, interaction volume, and apparatus construction are insufficient without effective anomaly prioritization and experimental reasoning.

The normal-control scene provides an additional test of evidential discipline. Across models, 36.1% of control reports are rejected because they assert altered laws or unsupported explanations under ordinary physics. One example attributes longer travel on a decorative gray tile to reduced friction, although the ball begins each comparison with different residual velocities. The control therefore exposes a recurring confound: models often convert a correlation in an uncontrolled comparison into a causal physical explanation.

From anomaly detection to research-question formation

The most important qualitative analysis concerns whether an explorer changes its plan after observing an unexpected event. Scene 15 contains a northward force acting on the red ball. Every nonzero push or throw restarts a 10-second pause, after which the force resumes. This law is difficult because the intervention both changes the ball’s immediate motion and modifies the persistence of the hidden force.

GPT-6-Astra notices that the red ball restarts after a delay and suspends its planned collision experiment. It measures the delay and tests whether pushing the blue ball produces the same reset. The resulting investigation identifies the relevant trigger and the temporary nature of the interruption. Gemini-3.7-Flash observes a similar restart but continues its original plan, leaving the anomaly unexplained.

Figure 3: GPT-6-Astra converts a delayed restart into a research question, whereas Gemini-3.7-Flash observes the same event without revising its plan.

This comparison isolates a capability that ordinary exploration metrics do not capture: adaptive allocation of experimental effort. The relevant behavior is not merely noticing an event in the observation stream, but treating it as evidence that the current model is inadequate and selecting an intervention that resolves the discrepancy.

Providing explicit guidance improves discovery from 16.7% to 28.6%, while evidence support remains at 97.6%. Mean exploration length falls from 39.3 to 30.9 turns. The implication is specific: guidance primarily improves the transition from anomaly recognition to law identification, rather than the ability to document experimental evidence once an investigation has begun. This intervention also confirms that the free-exploration result measures autonomous prioritization, not only raw physical reasoning.

Experimental discrimination and evidential validity

The framework requires explorers to distinguish competing explanations rather than merely accumulate observations. Scene 12 illustrates this requirement. A red-blue collision changes the balls’ colors, which is consistent with at least two hypotheses: color exchange between the objects, or an independent red-blue color-flipping rule. GPT-6-Astra tests a red-red collision. Color exchange predicts that both objects remain red, whereas independent flipping predicts that both become blue. The observed outcome supports exchange and rejects the alternative.

Figure 4: A discriminative experiment tests color exchange against independent color flipping by applying the intervention to two red objects.

This example shows why controlled interventions are necessary for law discovery. Repeated observations under the original configuration would not identify which explanation generalizes. The explorer must select a condition under which the hypotheses make different predictions.

The paper’s report protocol further constrains retrospective reasoning. Conclusions must cite turns from the exploration record, preserve uncertainty expressed during exploration, and avoid introducing explanations that were not previously recorded. This design addresses a common failure mode in agent evaluation: allowing a final report to reinterpret an exploratory trajectory as if the agent had planned its experiments in advance.

Apparatus construction as experimental design

LongChang allows the explorer to determine when primitive actions are insufficient and to construct a measurement device. In Scene 15, GPT-6-Astra builds a weighing platform to compare the loads associated with different balls. The initial platform is unstable because the red ball drifts off after release. The explorer redesigns it by adding retaining walls and increasing the platform’s vertical travel, then obtains stable readings.

Figure 5: GPT-6-Astra iteratively redesigns a weighing platform after diagnosing instability in the initial apparatus.

Access to apparatus increases discovery from 4.8% to 16.7% and law correctness from 4.8% to 19.0%, while evidence support remains perfect or near-perfect. This result supports the claim that construction can expand the set of experimentally accessible comparisons. It does not, however, establish that tool use itself caused the improvement: the explorer chooses whether to build, and the comparison changes both the available action space and the instructions associated with it. The paper correctly interprets the intervention as access to construction, not as a causal estimate of successful tool use.

The apparatus results also expose a distinction between engineering competence and scientific competence. Several models successfully construct devices but fail to formulate the experiments that would resolve the relevant law. Conversely, tool redesign is useful only when embedded in an investigative loop that diagnoses why the first design failed and connects the redesign to a hypothesis.

Resource scaling and early stopping

Increasing the turn budget from 100 to 500 does not improve discovery. GPT-6-Astra uses approximately 40 turns under either condition and stops early, leaving some hypotheses unresolved. Discovery declines slightly from 16.7% to 14.3%, while evidence support remains at 100% in the reported comparison.

This result is contradictory to a simple compute-scaling account of exploratory discovery. More available interaction time has little effect when the agent terminates before exhausting the budget. The bottleneck is therefore not only the number of available interventions, but the policy governing when to stop, when to revisit an anomaly, and when evidence is sufficient for a claim.

The apparatus intervention yields a larger improvement than the turn-budget intervention, suggesting that experimental expressivity can matter more than raw interaction count. Nevertheless, the study does not establish whether longer budgets would help under a stopping policy explicitly optimized for unresolved questions, because the experiments retain the original model’s termination behavior.

Scene evolution from corrected errors

LongChang uses exploration records not only for evaluation but also to generate more difficult environments. In Scene 15, the explorer initially interprets the temporary cessation of red’s motion as evidence that no persistent force exists. Later observations correct this interpretation: the force resumes after 10 seconds. The framework uses this corrected error to create three new scenes.

• Generation replaces the fixed northward force with mutual red-blue repulsion while retaining the pause.
• Refinement makes the force depend on blue’s position relative to a spatial square in addition to the pause.
• Composition combines the paused force with delayed velocity copying from red to green.

These transformations preserve the mechanism associated with the error while modifying its dependencies or combining it with another law. They therefore target unresolved distinctions rather than generating unrelated random tasks.

The results are negative but informative. GPT-6-Astra discovers the source laws at rates of 33.3% in Scene 15 and 66.7% in Scene 14, but achieves 0% discovery in all three evolved scenes. Evidence support remains 100%, meaning the model records genuine observations, but no report identifies the complete evolved law. In the generated scene it detects position-dependent repulsion without recovering the pause; in the refined scene it observes northward motion without linking it to blue’s location; and in the composed scene it records red’s drift and green’s intermittent motion without separating the 10-second and 12-second timing rules.

The result demonstrates that scene evolution successfully exposes missing conditions and dependencies. It also shows that a corrected error does not automatically yield a robust abstraction: the explorer can repair one interpretation while remaining unable to generalize it across changed triggers, spatial conditions, or compositional timing.

Limitations and open questions

The benchmark’s physical worlds are fully simulated, and the ground-truth laws are known to the evaluator. This enables precise scoring of correctness and evidence support, but it limits direct conclusions about scientific discovery in settings where the underlying answer is unknown. The paper explicitly notes that extending the framework to open scientific questions requires evaluation methods that do not presuppose a complete reference law.

The reported results also depend on LLM-based judges and scene designers. Although exploration records and simulator state provide structured evidence, automated evaluators may still misclassify correctness or support. The paper acknowledges the risk that an LLM judge could endorse unsupported findings and thereby guide subsequent scene evolution in the wrong direction.

The base suite contains one altered law per scene, whereas the evolved composition experiment introduces multiple interacting laws. It remains open how discovery rates change with systematic increases in law complexity, stochasticity, perceptual noise, partial observability, or ambiguous ground truth. It is also unresolved whether improved stopping policies, explicit uncertainty tracking, or hypothesis-management mechanisms would convert the additional turn budget into useful experimentation.

Finally, the model comparison is observational across systems with potentially different prompting, action reliability, and tool-use behavior. The controlled interventions isolate task guidance, turn budget, and apparatus access, but they do not identify which internal reasoning components distinguish GPT-6-Astra from the other models. In particular, the paper leaves open whether the advantage arises from anomaly detection, planning, causal abstraction, memory organization, or more reliable execution.

Conclusion

LongChang evaluates a demanding form of autonomous physical inquiry in which LLMs must choose what to investigate, construct experiments, discriminate among explanations, and state only claims supported by their records. Its central finding is that these capabilities remain sharply dissociated. GPT-6-Astra reaches 16.7% fully supported discovery, while other models either fail to identify the law, fail to support their identification, or ignore anomalies despite extensive interaction and tool construction.

The framework’s main methodological contribution is the joint evaluation of research-question formation, experimental discrimination, evidence-calibrated reporting, and scene evolution. Its results show that free exploration is substantially harder than task-directed law recovery, and that unresolved conditions—not merely missing observations—are the dominant source of failure.