Scientist World launches 1,000 executable AI research environments
Princeton University researchers introduced Scientist World, a system that converts research papers into executable environments where AI Scientist can propose methods, run experiments, and compare results with reproduced paper baselines. According to the paper, its first release contains 1,000 environments validated by independent agents and retains about 6,400 hours of research trajectories across models, reasoning settings, and two research modes. The trajectories include ideas, revisions, experiment logs, scores, and development histories. The team also trained Qwen3-8B on the trajectories, using scores from executed experiments as training rewards and a fixed GPT-5.6-Sol xhigh executor to evaluate proposed ideas on 10 held-out paper environments.
The same environments are designed for three uses: evaluating whether AI Scientist can advance research, providing trajectory data for training, and supplying an experimental test bed with measurable rewards. This makes Scientist World relevant to research on both AI research capability and model training, although the reported results do not establish that it eliminates data contamination or demonstrates a general, independently executable research ability.
AI Scientist 已经能够阅读科研文献,但要在已有研究的基础上继续探索,仅仅读懂论文还不够。人类科学家可以从论文中提炼问题、借鉴方法,再搭建实验检验新的想法,AI Scientist 同样需要能够实际运行实验、验证想法的研究环境。
这样的环境既要保留论文中的研究问题,也要提供可运行的方法和衡量改进的标准。然而,现有研究环境往往依赖人工挑选的任务、过去的竞赛题目或固定的模拟世界,难以同时做到大规模扩展、持续更新,并通过实际运行的计算实验检验研究成果。
为此,普林斯顿大学团队提出了 Scientist World,将研究论文转化为可执行的环境,让 AI Scientist 能在其中提出新方法、开展实验,再将实验结果与原论文的复现结果进行比较,衡量自身取得了多少改进。
#Princeton University#Scientist World#AI Scientist#Qwen3-8B#GPT-5.6-Sol xhigh#Frontier-Normalized Improvement