Code Generation on Python 10 sequential 20-step campaigns (end-of-chain)
25.9End pass@1Qwen2.5-7B + Best-ckpt oracle (peak-ckpt, no gate)
Evaluation Results
| Method | Links | |
|---|---|---|
| Qwen2.5-7B + Best-ckpt oracle (peak-ckpt, no gate)Scale=7B (ref.), n=12026.06 | 25.9 | |
| Qwen2.5-7B + CARE v2 (gate + revision)Scale=7B, n=52026.06 | 13.8 | |
| Qwen2.5-7B + Naive (end-ckpt)Scale=7B, n=52026.06 | 11.8 | |
| Qwen2.5-3B + CARE v2 (gate + revision)Scale=3B, n=52026.06 | 9.5 | |
| Qwen2.5-3B + Naive (end-ckpt)Scale=3B, n=52026.06 | 4.9 |