LongChang 评测 12 个 LLM 在 14 个模拟世界中的自主科学发现能力

AI 导读

LongChang（龙场）项目用 14 个物理规律被根本改变的模拟世界，评测 12 个 LLM 的自主科学发现能力。GPT-6-Astra 的发现率为 16.7%，其余 11 个模型均为 0%，将探索预算提高 5 倍也没有改善结果。项目提供交互式演示、arXiv 论文和代码。

正文

Can AI discover physical laws that humanity has never known?

Excited to share LongChang (龙场), our team's work evaluating frontier AI models on autonomous scientific discovery.

Named after the place where Chinese philosopher Wang Yangming (王阳明) attained enlightenment, LongChang challenges 12 LLMs to explore 14 simulated worlds with fundamentally altered physical laws.

📊 Results:

- GPT-6-Astra: 16.7% discovery rate

- All other 11 models: 0%

- 5× more exploration budget brought no improvement.

Can AI discover what it doesn't even know to ask?

🌐 Try our interactive demo directly on our website! We'd love to hear what you think.

📄 Paper: arxiv.org/abs/2610.37649

💻 Code: github.com/EnVision-Research…

first author @ZhenYang_CS