Tsinghua's AutoSFT uses multi-agent collaboration to automate SFT data research
TsinghuaNLP · hf · 2026-10-16
Tsinghua NLP, with BUPT, UIUC and collaborators, released AutoSFT, a multi-agent system that automates data research for supervised fine-tuning.
- Observes that a single agent given the full training stack converges on one action (adjusting category proportions) and its context fills with execution output mid-search
- Designs a full data action space — adjusting proportions, pruning, repairing, replacing, synthesizing — mirroring what human data engineers do
- Separating execution, analysis, and synthesis across agents improves evaluation scores over single-agent systems
Code is open-sourced on GitHub (Thinking-Space/AutoSFT).
More from coding & agent
- SpecAgent: small model drafts tool calls, large model verifies in parallel to cut latency — Zilin Wang · 2026-10-16
- Salesforce's CLIFT turns sparse judge feedback into rewards for open-source web agents — Salesforce · 2026-10-16
- Agent spots invalid SPF record and autonomously fixes it in Vercel DNS via browser — dkundel · 2026-10-16
- MCPJam engineer asks how to test if ChatGPT plugins actually see latest UI selection — Paoli99 · 2026-10-16
- ChatGPT revives unplayable 2000s game Bookworm on Mac and iPhone in 15 minutes — Shleemy_Pants · 2026-10-16
- Agentflow: open-source collective memory network for agents via FastMCP and pgvector — agentflow0328 · 2026-10-16