Can LLMs Uncover Hidden Physical Laws through Free Exploration?

Can LLMs Uncover Hidden Physical Laws through Free Exploration?. By Zhen Yang and three others, from the Hong Kong University of Science and Technology, Guangzhou, Alibaba Group, and two other institutions. Published on the twenty-seventh of October.

Scientific discovery often starts with something unexpected, an observation that raises a question nobody was asking. The authors want to know whether large language models can do that first step. Put one in a world where the physical laws have been altered, where heavier objects fall more slowly, say. Will it notice that something is off, and then work out the law through its own experiments, without anyone telling it what to look into? The question is withheld on purpose. Finding it is part of what is being tested, before any law can be worked out.

That is the ability being measured: unprompted discovery. The evaluated models act as explorers. They manipulate objects and build tools to chase their own questions, with no research task and no hints about which laws have changed. Success means arriving at a correct account of an altered law, backed by experiment.

The framework is called LongChang, and it runs automatically. It creates simulated worlds in which one or more laws differ from ordinary physics. Separate LLM agents generate scenes and review them, review the tools the explorers design, and evaluate the final reports, checking both correctness and experimental support against the altered laws and the exploration records. Those records also guide how the scenes evolve for further exploration, so the whole cycle runs with no human in it. The design has a loop at its heart. An explorer acts, the record of what it did shapes the next scene, and the next scene gives it a fresh chance to notice what it missed. Judging is done by agents that know the altered laws, while the explorer never does.

Twelve LLMs were evaluated. GPT-6-Astra reaches a discovery rate of sixteen point seven percent. The other eleven score zero percent. The exploration records show what the explorers missed: anomalies left uninvestigated, and altered laws identified only in part.

The result marks a clear boundary. Following a stated research question is one skill. Noticing that a question exists, in a world that quietly breaks the rules, is another, and on this evidence nearly every model tested does not yet have it.