Persona Alignment on Consistent-LLMs chatting domain
0.995Persona Consistency (PC)3-Step-PG
Evaluation Results
| Method | Links | |
|---|---|---|
| 3-Step-PGBackbone=Qwen, Evaluation Methodology=Step-level2026.03 | 0.995 | |
| 3-Step-PGBackbone=Qwen, Evaluation Methodology=Rollout2026.03 | 0.985 | |
| 5-Step-PGBackbone=Qwen, Evaluation Methodology=Step-level2026.03 | 0.964 | |
| 5-Step-PGBackbone=Qwen, Evaluation Methodology=Rollout2026.03 | 0.958 | |
| 2-Step-PGBackbone=Llama, Evaluation Methodology=Step-level2026.03 | 0.939 | |
| 2-Step-PGBackbone=Llama, Evaluation Methodology=Rollout2026.03 | 0.929 | |
| 4-Step-PGBackbone=Qwen, Evaluation Methodology=Step-level2026.03 | 0.925 | |
| GreedyPGBackbone=Qwen, Evaluation Methodology=Step-level2026.03 | 0.881 | |
| PGBackbone=Qwen, Evaluation Methodology=Rollout2026.03 | 0.867 | |
| GreedyPGBackbone=Llama, Evaluation Methodology=Step-level2026.03 | 0.85 | |
| PPOBackbone=Qwen, Evaluation Methodology=Step-level2026.03 | 0.845 | |
| 3-Step-PGBackbone=Llama, Evaluation Methodology=Step-level2026.03 | 0.835 | |
| 4-Step-PGBackbone=Qwen, Evaluation Methodology=Rollout2026.03 | 0.831 | |
| 5-Step-PGBackbone=Llama, Evaluation Methodology=Step-level2026.03 | 0.824 | |
| 3-Step-PGBackbone=Llama, Evaluation Methodology=Rollout2026.03 | 0.821 | |
| PPOBackbone=Qwen, Evaluation Methodology=Rollout2026.03 | 0.818 | |
| 2-Step-PGBackbone=Qwen, Evaluation Methodology=Step-level2026.03 | 0.813 | |
| PGBackbone=Qwen, Evaluation Methodology=Step-level2026.03 | 0.813 | |
| 5-Step-PGBackbone=Llama, Evaluation Methodology=Rollout2026.03 | 0.802 | |
| GreedyPGBackbone=Llama, Evaluation Methodology=Rollout2026.03 | 0.799 | |
| 2-Step-PGBackbone=Qwen, Evaluation Methodology=Rollout2026.03 | 0.782 | |
| BaseBackbone=Qwen, Evaluation Methodology=Step-level2026.03 | 0.775 | |
| PGBackbone=Llama, Evaluation Methodology=Step-level2026.03 | 0.771 | |
| GreedyPGBackbone=Qwen, Evaluation Methodology=Rollout2026.03 | 0.762 | |
| BaseBackbone=Qwen, Evaluation Methodology=Rollout2026.03 | 0.76 | |
| PGBackbone=Llama, Evaluation Methodology=Rollout2026.03 | 0.727 | |
| 4-Step-PGBackbone=Llama, Evaluation Methodology=Step-level2026.03 | 0.715 | |
| 4-Step-PGBackbone=Llama, Evaluation Methodology=Rollout2026.03 | 0.692 | |
| PPOBackbone=Llama, Evaluation Methodology=Step-level2026.03 | 0.558 | |
| PPOBackbone=Llama, Evaluation Methodology=Rollout2026.03 | 0.556 | |
| BaseBackbone=Llama, Evaluation Methodology=Rollout2026.03 | 0.478 | |
| BaseBackbone=Llama, Evaluation Methodology=Step-level2026.03 | 0.478 |