Multi-Agent Collaboration Role Overstepping on SWE easy (dev)
0.4Overstepping Rate (<INFO>)ChatDev
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| ChatDevCEO Configuration=FT, CPO Configuration=FT, Backbone=Llama, System=ChatDev, Role Clarity Regularization=enabled2026.04 | 0.4 | 0.4 | -44.8 | -46 | |
| ChatDevCEO Configuration=FT, CPO Configuration=Base, Backbone=Llama, System=ChatDev, Role Clarity Regularization=enabled2026.04 | 0.8 | 0.8 | -44.4 | -45.6 | |
| ChatDevCEO Configuration=Base, CPO Configuration=FT, Backbone=Llama, System=ChatDev, Role Clarity Regularization=enabled2026.04 | 16.8 | 16.8 | -28.4 | -29.6 | |
| ChatDevCEO Configuration=Base, CPO Configuration=Base, Backbone=Llama, System=ChatDev, Role Clarity Regularization=enabled2026.04 | 45.2 | 46.4 | — | — |