📄 中文摘要
随着人工智能代理的出现，自动科学发现已成为一个可行的目标。许多近期研究构建了能够进行机器学习研究的代理系统，但缺乏系统化的训练方法，现有的大型语言模型往往生成看似合理但无效的想法。为了解决这一问题，提出了一种新颖的合成环境生成管道，专门针对机器学习代理。该管道自动合成与SWE-agent框架兼容的机器学习挑战，涵盖主题采样、数据集提议和代码生成。生成的合成任务基于真实的机器学习数据集，确保提出的数据集具备实际应用价值。
通过合成任务扩展的人工智能科学家
出处: AI Scientist via Synthetic Task Scaling
发布: 2026年3月19日
📄 English Summary
AI Scientist via Synthetic Task Scaling
The emergence of AI agents has made automatic scientific discovery a feasible goal. Recent works have built agentic systems capable of conducting machine learning research but lack a principled approach to train these agents. Current large language models often produce plausible yet ineffective ideas. To address this, a novel synthetic environment generation pipeline is proposed, specifically targeting machine learning agents. This pipeline automatically synthesizes machine learning challenges compatible with the SWE-agent framework, encompassing topic sampling, dataset proposal, and code generation. The resulting synthetic tasks are grounded in real machine learning datasets, ensuring that the proposed datasets have practical applicability.