Logit Divergence Analysis on Physical reasoning domain (calibration set)
0.92Mean KL Divergence (nats)µCRASP
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| µCRASPBase Model=Qwen2.5-VL-7B, Pruning Ratio=30%2026.05 | 0.92 | 1 | |
| LLM-PrunerBase Model=Qwen2.5-VL-7B, Pruning Ratio=30%2026.05 | 1.53 | 1.7 | |
| Wanda-SPBase Model=Qwen2.5-VL-7B, Pruning Ratio=30%2026.05 | 3.31 | 3.6 | |
| FLAPBase Model=Qwen2.5-VL-7B, Pruning Ratio=30%2026.05 | 4.11 | 4.5 | |
| Attribution PruningBase Model=Qwen2.5-VL-7B, Pruning Ratio=30%2026.05 | 10.51 | 11.4 |