Stanford paper: self-organizing agent team hits 66.7%, beats best single model's 48.8%
dair_ai · x · 2026-09-23
A Stanford/Together AI paper shows why letting agent teams learn their own way of working together pays off:
- Three models (o3-mini, Claude Sonnet 4, DeepSeek-V3) as a self-organizing team averaged 66.7% across five math/physics benchmarks
- The strongest single member scored 48.8%; a perfect router over independent answers reached 59.0%
- On AIME 2026 the team hit 71.2%, 13.4 points above the router
Key mechanism: one member reviews past team exchanges and rewrites the teamwork strategy—roles, phase ordering, participation, answer merging. Strategies learned from just 15 AIME 2024 problems transferred unchanged to held-out problems and four new benchmarks.
More from coding & agent
- YouArt Desktop launches: AI agent edits video in CapCut, DaVinci and Premiere from client feedback — mhdfaran · 2026-11-11
- Boubou: one-command local voice AI that serves your phone and robots over Wi-Fi — LuozhuZhang · 2026-11-11
- Python 3.15 makes AttributeError actionable: suggests .append for .push mix-ups — KhuyenTran16 · 2026-11-11
- sentence-transformers maintainer releases new embedding pooling training method, code & checkpoints — tomaarsen · 2026-11-11
- droid-mcp turns your Android phone into an MCP server with 160+ tools for Claude Code — zvone1122 · 2026-11-11
- The decade of agents: simulate real customers to test your agent before launch — EXM7777 · 2026-11-11