VPRune: Efficient Training-free Pre-LLM Visual Token Pruning
VPRune cuts visual tokens before they reach the language model without requiring new training.
The paper says performance drops from heavy visual-token pruning come from text-guided selection bias, lost information, and distorted positions after compaction. VPRune addresses those with visual-only diversity selection, similarity-guided token recycling, and position-preserving restoration. On FastVLM-1.5B, the authors report a better accuracy-compression trade-off across multiple vision-language benchmarks, especially under aggressive compression. They also say edge-device tests show lower end-to-end latency while preserving stronger task performance. ArXiv · AI/CL/LG's note
The paper says performance drops from heavy visual-token pruning come from text-guided selection bias, lost information, and distorted positions after compaction. VPRune addresses those with visual-only diversity selection, similarity-guided token recycling, and position-preserving restoration. On FastVLM-1.5B, the authors report a better accuracy-compression trade-off across multiple vision-language benchmarks, especially under aggressive compression. They also say edge-device tests show lower end-to-end latency while preserving stronger task performance. ArXiv · AI/CL/LG's note
score 5