Code Generation on CodeContests EASY difficulty (held-out)
34Pass@1Agon
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| AgonBackbone=Qwen3-1.7B, Training=single training run2026.07 | 34 | 3,900 | |
| Cooperative exchangeBackbone=Qwen3-1.7B, Training=single training run2026.07 | 29 | 4,800 | |
| Vanilla GRPOBackbone=Qwen3-1.7B, Training=single training run2026.07 | 24 | 7,300 | |
| Zero-shotBackbone=Qwen3-1.7B, Evaluation=single evaluation2026.07 | 18 | 5,200 |