Multi-modal Reasoning on MMStar
63.78AccuracySINKTRACK
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SINKTRACKBase LLM=Qwen2.5-VL-7B-Instruct2026.04 | 63.78 | 77.86 | |
| SINKTRACKBase LLM=Gemma3-12B-Instruct2026.04 | 61.6 | 75.81 | |
| CoTBase LLM=Gemma3-12B-Instruct2026.04 | 60.56 | 75.09 | |
| VanillaToken Reduction=0% (Upper Bound)2026.04 | 57.1 | — | |
| CoTBase LLM=Qwen2.5-VL-7B-Instruct2026.04 | 55.84 | 71.29 | |
| DASHToken Reduction=75.00%2026.04 | 55.4 | — | |
| SINKTRACKBase LLM=Gemma3-4B-Instruct2026.04 | 53.04 | 69.11 | |
| CoTBase LLM=Gemma3-4B-Instruct2026.04 | 51.36 | 67.52 | |
| PruMerge+Token Reduction=75.00%2026.04 | 48.1 | — | |
| SINKTRACKBase LLM=Qwen2.5-VL-3B-Instruct2026.04 | 47.24 | 61.38 | |
| VisionZipToken Reduction=75.00%2026.04 | 47.1 | — | |
| DARTToken Reduction=75.00%2026.04 | 46.9 | — | |
| FastVToken Reduction=75.00%2026.04 | 44.8 | — | |
| DirectBase LLM=Gemma3-12B-Instruct2026.04 | 44.58 | 60.89 | |
| DASHToken Reduction=88.89%2026.04 | 43.4 | — | |
| DASHToken Reduction=93.75%2026.04 | 40.4 | — | |
| VisionZipToken Reduction=88.89%2026.04 | 40.3 | — | |
| FastVToken Reduction=88.89%2026.04 | 40.2 | — | |
| DARTToken Reduction=88.89%2026.04 | 39.4 | — | |
| PruMerge+Token Reduction=88.89%2026.04 | 38.5 | — | |
| FastVToken Reduction=93.75%2026.04 | 37.4 | — | |
| DirectBase LLM=Qwen2.5-VL-7B-Instruct2026.04 | 37.29 | 54.24 | |
| DirectBase LLM=Gemma3-4B-Instruct2026.04 | 35.47 | 52.17 | |
| VisionZipToken Reduction=93.75%2026.04 | 35.3 | — | |
| CoTBase LLM=Qwen2.5-VL-3B-Instruct2026.04 | 35.24 | 50.68 | |
| PruMerge+Token Reduction=93.75%2026.04 | 33.4 | — | |
| DARTToken Reduction=93.75%2026.04 | 33.2 | — | |
| DirectBase LLM=Qwen2.5-VL-3B-Instruct2026.04 | 24.89 | 39.54 |