Scientists have long relied on unexpected observations to spark new discoveries in physics, but what if artificial intelligence could independently notice strange phenomena and figure out new laws of nature? A newly published research paper explores whether large language models (LLMs)—advanced AI systems trained to understand and generate human language—can autonomously explore simulated worlds with altered physical rules and uncover the hidden laws governing them. This study is important because it tests AI’s ability to perform scientific discovery without explicit instructions, potentially paving the way for automated research assistants in the future.
Key Takeaways
- The researchers developed LongChang, a framework that creates virtual worlds where physical laws differ from our real-world physics.
- LLMs acted as explorers, manipulating objects and building tools to investigate anomalies without being told what to look for.
- Out of 12 tested LLMs, only GPT-6-Astra managed to discover altered physical laws, with a discovery success rate of 16.7%.
- The exploration process revealed some uninvestigated anomalies, indicating that even the best models only partially identified the new laws.
To test AI’s ability to independently discover new physics, the researchers designed LongChang, an automated system that simulates worlds where one or more physical laws differ from those we know. For example, in one world, heavier objects might fall more slowly—contradicting real-world gravity. The LLMs were placed as virtual explorers in these worlds, with the freedom to interact with objects, conduct experiments, and create tools to aid their investigations. Importantly, the models were not given any specific research questions or hints about the altered laws—they had to notice the unexpected behaviors on their own and try to explain them.
The system works through a multi-agent setup: separate LLMs generate scenes, design tools, and review experimental reports. These agents evaluate the correctness of findings based on the altered physics and keep track of exploration records, which then guide further experiments. This creates a feedback loop allowing the AI to conduct a full cycle of scientific inquiry—observation, hypothesis, experimentation, and evaluation—without human intervention.
The results showed that most LLMs struggled with this open-ended discovery task. Only GPT-6-Astra was able to identify the new physical laws in some cases, achieving a discovery rate of about 17%. The other 11 models failed to detect the altered physics. Even GPT-6-Astra’s success was partial, with some anomalies remaining unexplored or only partially explained. This indicates that while LLMs have potential to autonomously uncover new scientific principles, their current capabilities are limited and require further development.
These findings highlight both the promise and challenges of using AI for autonomous scientific discovery. If improved, such systems could assist researchers by generating hypotheses and designing experiments in complex fields where unexpected phenomena arise. Future work may focus on enhancing LLMs’ reasoning abilities, incorporating more sophisticated experimentation tools, and expanding the diversity of simulated environments. While fully independent AI scientists are still a goal for the future, this study provides an important step toward machines that can explore and learn about the physical world on their own.
Based on research published on arXiv by Zhen Yang, Wen Wang, Xander Xu et al..