Chatbot Evaluation on MT-Bench High-Disagreement (Top 20%)
8.72Human ScorerDPO + DARC-ϵ
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| rDPO + DARC-ϵdecoding_strategy=DARC-ϵ2026.03 | 8.72 | 0.67 | 7.38 | 7.48 | |
| DARC-ϵ2026.03 | 8.34 | 0.65 | 7.34 | 7.6 | |
| cDPO + DARC-ϵdecoding_strategy=DARC-ϵ2026.03 | 8.31 | 0.63 | 7.05 | 7.54 | |
| rDPOdecoding_strategy=Best-of-K2026.03 | 8.25 | 0.91 | 6.43 | 7.02 | |
| DARCversion=base2026.03 | 8.18 | 0.74 | 7 | 7.43 | |
| DARC-τ2026.03 | 8.11 | 0.65 | 6.81 | 7.35 | |
| DeAL2026.03 | 7.99 | 1.13 | 5.73 | 7.33 | |
| cDPOdecoding_strategy=Best-of-K2026.03 | 7.96 | 0.89 | 6.18 | 7.15 | |
| RBoN2026.03 | 7.95 | 0.81 | 6.33 | 7.29 | |
| MC-Dropout2026.03 | 7.88 | 0.84 | 6.2 | 7.25 | |
| BoPvariant=HedgeTune2026.03 | 7.83 | 0.87 | 6.09 | 6.81 | |
| Caution2026.03 | 7.81 | 0.85 | 6.11 | 6.94 | |
| Basedecoding_strategy=Best-of-K2026.03 | 7.62 | 1 | 5.62 | 6.1 |