GEA treats agent groups as the evolution unit, hitting 71% SWE-bench Verified with zero human help
xwang_lk · x · 2026-10-07
The COLM 2026 paper Group-Evolving Agents (GEA) proposes a new open-ended self-improvement paradigm: instead of tree-structured single-agent evolution, a group of agents explicitly shares experience and learned artifacts during evolution, avoiding diversity trapped in isolated branches. GEA reaches 71.0% on SWE-bench Verified and 88.3% on Polyglot with zero human intervention, beating prior self-evolving methods and matching or exceeding top human-designed agent frameworks (arXiv:2602.04837).
Related event: GEA: Group-Evolving Agents Achieve SWE-bench 71% at COLM 2026(3 posts)→
More from coding & agent
- Claude Opus rebuilt Terraria in 3D: 346K lines of C# in 8 days, zero hand-written code — Promptmethus · 2026-10-07
- Dev puts the entire game of Minecraft inside Twitter with working multiplayer — jaivinwylde · 2026-10-07
- Scikit-learn Team Launches Skore, an Enterprise Platform for Tabular AI Built With AI Agents — GaelVaroquaux · 2026-10-07
- Amateur Team Discloses Identity-Constitutive Memory Protocol for AI Agents, Keeps Core Docs Secret — AdeptBathroom3318 · 2026-10-07
- Prompting agents to mimic Modal's GPU glossary styling yields polished HTML artifacts — sumitdotml · 2026-10-07
- Jordan and coauthors tackle when to stop generator-verifier loops while controlling false discovery — _onionesque · 2026-10-07