Can LLMs Uncover Hidden Physical Laws through Free Exploration?

Can LLMs Uncover Hidden Physical Laws through Free Exploration?

Zhen Yang, Wen Wang, Xander Xu and Ying-Cong Chen introduce LongChang, a framework for testing whether LLMs can investigate unexpected behavior and uncover altered physical laws without being told what to investigate. LongChang creates simulated worlds where one or more laws differ from ordinary physics. The LLMs explore by manipulating objects and building tools, while separate agents generate and review scenes, tools and reports. The authors report that GPT-6-Astra achieved a discovery rate of 16.7%; the other 11 evaluated LLMs scored 0%. They also report that exploration records contained anomalies the LLMs had not investigated and only partial identification of the altered laws. The findings describe both discoveries and limits in this setup: exploration yielded discoveries, but clues went unpursued and the altered laws were only partly identified.

Publication: 26546021-f224-4293-87c3-6dacfe0f824a

attribution — pass: The draft names the four listed authors, presents LongChang as their framework and attributes the results with phrasing such as 'The authors report.' It does not present the source's claims as established fact or as Escape Meridian's view.

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

summary — pass: The reading leads with the actors, the framework and the open-ended setup, then gives the headline result and the stated limits. The question about what counted as a discovery is a real gap in the abstract. Optional repairs: spell out large language models once, add the abstract's 'heavier objects falling more slowly' example, and say that LLM agents judged report correctness. The closing line 'exploration yielded discoveries' is accurate, but it should keep in view that only one of 12 models scored above 0%.

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?

date — pass: evidence.originalPublished is 2026-10-27. This matches the arXiv v1 submission date. The feed date of 2026-10-29 is later and is correctly not used.

[Submitted on 27 Oct 2026]

support — pass: Each claim matches the abstract verbatim in substance: the framework, the explorer role, the 16.7% versus 0% result and the uninvestigated anomalies. The limitations stay within the evidence and do not turn a simulated benchmark score into real-world scientific ability. No benefits, costs or forecasts are invented.

Among 12 LLMs, GPT-6-Astra achieves a discovery rate of 16.7%, while the other 11 score 0%.

media — pass: The selected media ID is the verified direct link to the original arXiv abstract page, which is the source itself. No better source media was found.