Expert-Iteration RLVR on MedQA, HEAD-QA, ARC-C, and CaseHOLD
4Pathwise Clean ScoreCSA
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| CSAreplications per cell=20, alpha=pivotal alpha*2026.05 | 4 | 4 | 4 | 52 | |
| CRCreplications per cell=20, alpha=pivotal alpha*2026.05 | 4 | 0 | 0 | 2.8 | |
| LTTreplications per cell=20, alpha=pivotal alpha*2026.05 | 4 | 1 | 1 | 14.2 | |
| ConfFactreplications per cell=20, alpha=pivotal alpha*2026.05 | 3 | 1 | 1 | 25.5 | |
| Fixed-Thresholdreplications per cell=20, alpha=pivotal alpha*2026.05 | 2 | 4 | 2 | 66.1 | |
| NEX-Confreplications per cell=20, alpha=pivotal alpha*2026.05 | 1 | 4 | 1 | 79.8 | |
| ACIreplications per cell=20, alpha=pivotal alpha*2026.05 | 1 | 4 | 1 | 90.7 | |
| SAOCPreplications per cell=20, alpha=pivotal alpha*2026.05 | 0 | 4 | 0 | 90.3 | |
| Naive-Tuningreplications per cell=20, alpha=pivotal alpha*2026.05 | 0 | 4 | 0 | 87.2 | |
| Always-Actreplications per cell=20, alpha=pivotal alpha*2026.05 | 0 | 4 | 0 | 100 |