Reason scoring on human labels (held-out)
59wf1SFT+GRPO
Evaluation results
| Method | Links | ||
|---|---|---|---|
| SFT+GRPOEval mode=think, n=2542026.07 | 59 | 3.1 | |
| SFTEval mode=no-think, n=2442026.07 | 57.7 | — | |
| SFT+GRPOEval mode=no-think, n=2542026.07 | 55.9 | — | |
| BaseEval mode=no-think, n=2542026.07 | 52.9 | — | |
| BaseEval mode=think, n=2542026.07 | 44.9 | 8 | |
| Random baselineEval mode=–, n=2592026.07 | 43.6 | — | |
| Kimi-K2.6Eval mode=think, n=2442026.07 | 38.4 | — | |
| Kimi-K2.6Eval mode=no-think, n=2502026.07 | 36.8 | — |