Rationale Faithfulness Evaluation on PVP
99.5R-D ConsistencyQwen2.5-VL-7B-Instruct (Reasoning-SFT)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Qwen2.5-VL-7B-Instruct (Reasoning-SFT)Student Model=Qwen2.5-VL-7B-Instruct, Setup=Reasoning-SFT2026.05 | 99.5 | 84.7 | 74.6 | |
| Phi-3.5-Vision-Instruct (Base)Student Model=Phi-3.5-Vision-Instruct, Setup=Base2026.05 | 99 | 70.8 | 76.6 | |
| Phi-3.5-Vision-Instruct (Reasoning-SFT)Student Model=Phi-3.5-Vision-Instruct, Setup=Reasoning-SFT2026.05 | 99 | 70.3 | 66.5 | |
| Qwen2.5-VL-7B-Instruct (Base)Student Model=Qwen2.5-VL-7B-Instruct, Setup=Base2026.05 | 96.2 | 72.7 | 88.5 | |
| Phi-3.5-Vision-Instruct (GRPO)Student Model=Phi-3.5-Vision-Instruct, Setup=GRPO2026.05 | 90.4 | 75.1 | 51.7 | |
| Qwen2.5-VL-7B-Instruct (GRPO)Student Model=Qwen2.5-VL-7B-Instruct, Setup=GRPO2026.05 | 90 | 78 | 73.7 |