LongChang tests whether LLMs can discover altered physical laws
Original title: Can LLMs Uncover Hidden Physical Laws through Free Exploration?
Among 12 models, only GPT-6-Astra discovered laws, at a reported rate of 16.7 percent.
The paper introduces LongChang, an automated framework that builds simulated worlds whose physics differ from ordinary laws and lets LLMs explore without being told what to investigate. Separate agents generate scenes, review tool designs, and score reports using the altered laws and exploration records. The authors say GPT-6-Astra reached a 16.7 percent discovery rate while the other 11 models scored 0 percent. Records showed uninvestigated anomalies and only partial identification of the changed laws.
Unconfirmed: The 16.7 percent discovery rate is the authors' reported result on their simulated setup.
- Explorers manipulate objects and build tools without research hints.
- The cycle of scene evolution and review runs without human intervention.
- Eleven of the twelve evaluated LLMs scored a 0 percent discovery rate.
- Records showed only partial identification of the altered laws.
- Products
- LongChang
- AI models
- GPT-6-Astra
From the original article
1 min read at huggingface.co
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?
By Zhen Yang, Wen Wang, Xander Xu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.