AutoSFT：多智能体自动完成 SFT 数据研究全流程
清华系团队把 SFT 数据研究交给一组智能体自动跑，30 轮 IFEval 就超过单智能体跑 50 轮的成绩，论文和代码都开源了，搞微调的可以看看。
OpenBMB 与 THUNLP、BUPT、ModelBest、UIUC、CUHK 合作推出 AutoSFT，把 SFT 数据研究拆成 Executor、4 个分析子智能体（训练数据/失败案例/学习动态/研究历史）、带校验器的 Synthesis 写手和一个选下一配方的 meta 智能体。单智能体基线跑 33 轮只会调混合比例，上下文占用达 94.7% 且几乎没有分数提升。AutoSFT 在 30 轮内覆盖全部 6 类数据操作，IFEval 第 6 轮达 63.33，超过基线 50 轮的最佳成绩，同预算下 +6.5，第 30 轮达 67.10。Synthesis 单独使用时 IFEval +3.4、IFBench +2.9、MBPP+ 编码 +3.7。
🤖 What happens when you hand an agent the full SFT data-research task?
It only mixes ratios. No repaired samples, no new responses, no synthetic data — a word cloud full of "weights" and "proportions", and almost no score gain over dozens of rounds.
AutoSFT (THUNLP × BUPT × ModelBest × UIUC × CUHK) splits the data-research lifecycle into roles: a dedicated Executor, 4 Analysis Subagents (training data / failure cases / learning dynamics / research history), a Synthesis writer with a verifier, and a meta agent that picks the next recipe.
1️⃣ Single-agent baseline: 33 rounds, only Mixture; context hit 94.7%, compressed 10 times; Executor cuts compression 10→5 and shell events 58%→29%.
2️⃣ AutoSFT: all 6 action families appear across 30 rounds; IFEval round 6 = 63.33, beating the baseline's best over 50 rounds; +6.5 at the same budget; round 30 = 67.10.
3️⃣ Synthesis alone: IFEval +3.4, IFBench +2.9; MBPP+ coding +3.7. The takeaway: automating data research may depend more on covering diverse data actions than on a stronger single action.
📄 Paper: https://t.co/dzsl4fZe2J 💻 Code: https://t.co/fUphrIzn9M
#AIAgent #SFT #LLM #DataCentricAI #PostTraining #RSI #AgenticAI #THUNLP #AutoSFT #FineTuning
- arXiv cs.LG11-04 12:01原文
