Ranking on UltraFeedback clean (held-out)
0.362Kendall’s TauNominal PL
Evaluation Results
| Method | Links | |
|---|---|---|
| Nominal PLModel=Qwen3-8B, Noise condition=Clean2026.07 | 0.362 | |
| Robust PLModel=Qwen3-8B, Noise condition=near_tie, epsilon=0.4, rho=0.052026.07 | 0.356 | |
| Nominal PLModel=Qwen3-8B, Noise condition=near_tie, epsilon=0.42026.07 | 0.348 | |
| Robust PLModel=Qwen3-8B, Noise condition=Clean, rho=0.052026.07 | 0.347 | |
| Robust PLModel=Qwen3-8B, Noise condition=Clean, rho=0.102026.07 | 0.34 | |
| TV-DR-DPOModel=Qwen3-8B, Noise condition=Clean, rho=0.102026.07 | 0.338 | |
| Robust PLModel=Qwen3-8B, Noise condition=near_tie, epsilon=0.4, rho=0.102026.07 | 0.333 | |
| KLDPOModel=Qwen3-8B, Noise condition=Clean, tau=1.002026.07 | 0.331 | |
| Robust PLModel=Qwen3-8B, Noise condition=top_rank, epsilon=0.4, rho=0.052026.07 | 0.329 | |
| Robust PLModel=Qwen3-8B, Noise condition=near_tie, epsilon=1.0, rho=0.052026.07 | 0.327 | |
| Nominal PLModel=Qwen3-8B, Noise condition=near_tie, epsilon=1.02026.07 | 0.322 | |
| Nominal PLModel=Qwen3-8B, Noise condition=top_rank, epsilon=0.42026.07 | 0.321 | |
| TV-DR-DPOModel=Qwen3-8B, Noise condition=near_tie, epsilon=0.4, rho=0.102026.07 | 0.319 | |
| Robust PLModel=Qwen3-8B, Noise condition=near_tie, epsilon=1.0, rho=0.102026.07 | 0.319 | |
| KLDPOModel=Qwen3-8B, Noise condition=near_tie, epsilon=0.4, tau=1.002026.07 | 0.318 | |
| Nominal BT (DPO)Model=Qwen3-8B, Noise condition=Clean2026.07 | 0.316 | |
| TV-DR-DPOModel=Qwen3-8B, Noise condition=near_tie, epsilon=1.0, rho=0.102026.07 | 0.31 | |
| KLDPOModel=Qwen3-8B, Noise condition=near_tie, epsilon=1.0, tau=1.002026.07 | 0.31 | |
| Nominal BT (DPO)Model=Qwen3-8B, Noise condition=top_rank, epsilon=0.42026.07 | 0.302 | |
| Nominal BT (DPO)Model=Qwen3-0.6B, Noise condition=Clean2026.07 | 0.298 | |
| KLDPOModel=Qwen3-8B, Noise condition=top_rank, epsilon=0.4, tau=1.002026.07 | 0.296 | |
| Robust PLModel=Qwen3-8B, Noise condition=top_rank, epsilon=0.4, rho=0.102026.07 | 0.293 | |
| Nominal BT (DPO)Model=Qwen3-8B, Noise condition=near_tie, epsilon=0.42026.07 | 0.291 | |
| Nominal PLModel=Qwen3-0.6B, Noise condition=Clean2026.07 | 0.288 | |
| Robust PLModel=Qwen3-0.6B, Noise condition=Clean, rho=0.102026.07 | 0.284 | |
| TV-DR-DPOModel=Qwen3-0.6B, Noise condition=Clean, rho=0.102026.07 | 0.282 | |
| TV-DR-DPOModel=Qwen3-8B, Noise condition=top_rank, epsilon=0.4, rho=0.102026.07 | 0.282 | |
| Nominal BT (DPO)Model=Qwen3-8B, Noise condition=near_tie, epsilon=1.02026.07 | 0.281 | |
| KLDPOModel=Qwen3-0.6B, Noise condition=Clean, tau=1.002026.07 | 0.277 | |
| Robust PLModel=Qwen3-0.6B, Noise condition=Clean, rho=0.052026.07 | 0.276 | |
| Robust PLModel=Qwen3-0.6B, Noise condition=near_tie, epsilon=0.4, rho=0.052026.07 | 0.274 | |
| TV-DR-DPOModel=Qwen3-0.6B, Noise condition=near_tie, epsilon=0.4, rho=0.102026.07 | 0.27 | |
| Nominal PLModel=Qwen3-0.6B, Noise condition=near_tie, epsilon=0.42026.07 | 0.268 | |
| KLDPOModel=Qwen3-0.6B, Noise condition=near_tie, epsilon=0.4, tau=1.002026.07 | 0.267 | |
| Nominal BT (DPO)Model=Qwen3-0.6B, Noise condition=near_tie, epsilon=0.42026.07 | 0.266 | |
| Robust PLModel=Qwen3-0.6B, Noise condition=near_tie, epsilon=0.4, rho=0.102026.07 | 0.264 | |
| KLDPOModel=Qwen3-0.6B, Noise condition=top_rank, epsilon=0.4, tau=1.002026.07 | 0.263 | |
| Robust PLModel=Qwen3-0.6B, Noise condition=near_tie, epsilon=1.0, rho=0.052026.07 | 0.262 | |
| Nominal BT (DPO)Model=Qwen3-0.6B, Noise condition=near_tie, epsilon=1.02026.07 | 0.261 | |
| Robust PLModel=Qwen3-0.6B, Noise condition=top_rank, epsilon=0.4, rho=0.052026.07 | 0.251 | |
| KLDPOModel=Qwen3-0.6B, Noise condition=near_tie, epsilon=1.0, tau=1.002026.07 | 0.246 | |
| Robust PLModel=Qwen3-0.6B, Noise condition=near_tie, epsilon=1.0, rho=0.102026.07 | 0.246 | |
| TV-DR-DPOModel=Qwen3-0.6B, Noise condition=near_tie, epsilon=1.0, rho=0.102026.07 | 0.244 | |
| Nominal PLModel=Qwen3-0.6B, Noise condition=near_tie, epsilon=1.02026.07 | 0.244 | |
| Nominal BT (DPO)Model=Qwen3-0.6B, Noise condition=top_rank, epsilon=0.42026.07 | 0.244 | |
| Nominal PLModel=Qwen3-0.6B, Noise condition=top_rank, epsilon=0.42026.07 | 0.243 | |
| Robust PLModel=Qwen3-0.6B, Noise condition=top_rank, epsilon=0.4, rho=0.102026.07 | 0.236 | |
| TV-DR-DPOModel=Qwen3-0.6B, Noise condition=top_rank, epsilon=0.4, rho=0.102026.07 | 0.229 | |
| Robust PLModel=Qwen3-0.6B, Noise condition=top_rank, epsilon=1.0, rho=0.052026.07 | 0.154 | |
| Robust PLModel=Qwen3-8B, Noise condition=top_rank, epsilon=1.0, rho=0.052026.07 | 0.146 | |
| Nominal PLModel=Qwen3-0.6B, Noise condition=top_rank, epsilon=1.02026.07 | 0.119 | |
| Robust PLModel=Qwen3-0.6B, Noise condition=top_rank, epsilon=1.0, rho=0.102026.07 | 0.116 | |
| Nominal PLModel=Qwen3-8B, Noise condition=top_rank, epsilon=1.02026.07 | 0.103 | |
| Robust PLModel=Qwen3-8B, Noise condition=top_rank, epsilon=1.0, rho=0.102026.07 | 0.076 | |
| Nominal BT (DPO)Model=Qwen3-8B, Noise condition=top_rank, epsilon=1.02026.07 | 0.039 | |
| Nominal BT (DPO)Model=Qwen3-0.6B, Noise condition=top_rank, epsilon=1.02026.07 | 0.036 | |
| KLDPOModel=Qwen3-0.6B, Noise condition=top_rank, epsilon=1.0, tau=1.002026.07 | 0.031 | |
| TV-DR-DPOModel=Qwen3-0.6B, Noise condition=top_rank, epsilon=1.0, rho=0.102026.07 | 0.024 | |
| KLDPOModel=Qwen3-8B, Noise condition=top_rank, epsilon=1.0, tau=1.002026.07 | -0.016 | |
| TV-DR-DPOModel=Qwen3-8B, Noise condition=top_rank, epsilon=1.0, rho=0.102026.07 | -0.056 |