Multi-hop Question Answering on HotpotQA (val) (Token F1, Repair Metrics)
75Initial Token-level F1DISC
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| DISCModel=GPT-4o, K=32026.06 | 75 | 77.9 | 348 | 109 | 3.2 | 11.1 | |
| CoVeModel=GPT-4o2026.06 | 75 | 73.9 | 143 | 159 | 0.9 | 4.6 | |
| Self-RefineModel=GPT-4o2026.06 | 75 | 74.6 | 6 | 35 | 0.17 | 0.2 | |
| DISCModel=Claude Sonnet 4.5, K=32026.06 | 65.1 | 80.4 | 1,487 | 109 | 13.6 | 36.8 | |
| CoVeModel=Claude Sonnet 4.52026.06 | 65.1 | 80 | 1,687 | 224 | 7.5 | 41.8 | |
| Self-RefineModel=Claude Sonnet 4.52026.06 | 65.1 | 77.8 | 1,085 | 28 | 38.8 | 26.9 |