Hugging Face 这个开源智能体把 ML 研究全流程自动化了，做实验的团队可以直接用它跑论文复现和模型训练，省掉大量手动调参和写脚本的时间。
Hugging Face 开源了一个 CLI 智能体，能自动执行机器学习工程师的完整研究循环：从阅读论文、追踪引用，到拉取数据集、编写训练脚本，再到启动 GPU 任务并迭代优化。在科学推理任务中，它训练 Qwen3-1.7B 将 GPQA 得分从 10% 提升至 32%，超越 Claude Code 的 22.99%。在医疗领域，它因公开数据集质量不足，自动生成 1100 个合成样本并上采样 50 倍，在 HealthBench 上比 Codex 高出 60%。该智能体还具备自主诊断失败评估、恢复训练的能力，以 CLI 和 Web 应用形式提供，甚至可在手机上运行。
Hugging Face open-sourced a CLI agent that runs fu…
Hugging Face open-sourced a CLI agent that runs full research loop of an ML engineer.
You give it a prompt. It reads papers and walks citations. Then it pulls datasets and writes training scripts.
It launches GPU jobs and iterates until something works. On scientific reasoning, it trained Qwen3-1.7B and pushed GPQA from 10% to 32%.
Claude Code only reached 22.99%. For healthcare, it judged public datasets too weak.
So it generated 1,100 synthetic samples and upsampled 50x.
It beat Codex on HealthBench by 60%. On competitive math, rewards collapsed mid-run.
The agent ran ablations and recovered the training. Three things make it work: 1. Reads papers and citation graphs 2. Reformats messy data before training 3. Diagnoses failed evals autonomously Ships as a CLI and a web app you can run from your phone.