Role-Playing on RPGBench Character Shift (Generalization)
-0.8Deviation Score (Literature)RFT
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| RFTBackbone=Qwen3-4B2025.12 | -0.8 | -0.78 | -0.828 | -0.812 | |
| ICLBackbone=Meta-Llama3-8B2025.12 | -0.579 | -0.479 | -0.519 | -0.586 | |
| DPOBackbone=Meta-Llama3-8B2025.12 | -0.564 | -0.488 | -0.586 | -0.631 | |
| ICLBackbone=Qwen3-4B2025.12 | -0.372 | -0.282 | -0.321 | -0.349 | |
| DPOBackbone=Qwen3-4B2025.12 | -0.34 | -0.312 | -0.355 | -0.354 | |
| ThinkingRLBackbone=Meta-Llama3-8B2025.12 | -0.254 | -0.141 | -0.222 | -0.245 | |
| ThinkingRLBackbone=Qwen3-4B2025.12 | -0.198 | -0.137 | -0.191 | -0.186 | |
| RFTBackbone=Meta-Llama3-8B2025.12 | -0.18 | -0.184 | -0.193 | -0.186 | |
| ThinkingSFTBackbone=Qwen3-4B2025.12 | -0.131 | -0.093 | -0.141 | -0.128 | |
| SFTBackbone=Qwen3-4B2025.12 | -0.111 | -0.062 | -0.078 | -0.079 | |
| ThinkingSFTBackbone=Meta-Llama3-8B2025.12 | -0.104 | -0.05 | -0.08 | -0.11 | |
| DABackbone=Qwen3-4B2025.12 | -0.102 | -0.072 | -0.089 | -0.094 | |
| SFTBackbone=Meta-Llama3-8B2025.12 | -0.082 | -0.026 | -0.072 | -0.056 | |
| RLBackbone=Qwen3-4B2025.12 | -0.078 | -0.021 | -0.073 | -0.051 | |
| DABackbone=Meta-Llama3-8B2025.12 | -0.061 | -0.033 | -0.069 | -0.077 | |
| RLBackbone=Meta-Llama3-8B2025.12 | -0.056 | -0.008 | -0.076 | -0.042 | |
| CoRLBackbone=Qwen3-4B2025.12 | -0.049 | -0.04 | -0.038 | -0.035 | |
| CoRLBackbone=Meta-Llama3-8B2025.12 | -0.031 | -0.007 | -0.021 | -0.029 |