analiza aiqbench
VPRune addresses the performance degradation inherent in aggressive visual token pruning by introducing a training-free framework that preserves positional information and recycles similar tokens. This approach significantly reduces inference latency on edge devices while maintaining high accuracy across multiple vision-language benchmarks, offering a practical solution for deploying resource-constrained Large Vision-Language Models.
VPRune: Efficient Training-free Pre-LLM Visual Token Pruning
arXiv:2609.24485v2 Announce Type: replace-cross Abstract: Visual token pruning is a promising approach to reducing the inference cost of large vision-language models (LVLMs), yet aggressive token reduction often causes substantial performance degradation. We identify three key factors behind this degradation: text-guided selection bias, information loss from discarded tokens, and positional distortion caused by sequence compaction. Based on these observations, we propose \textbf{VPRune}, a training-free pre-LLM pruning framework consisting of visual-only diversity selection, similarity-guided token recycling, and position-preserving restoration. Experiments on FastVLM-1.5B across multiple vision-language benchmarks demonstrate that VPRune achieves a favorable accuracy--compression trade-off, with particularly pronounced advantages under aggressive compression. Furthermore, evaluations on edge-device show that VPRune effectively reduces end-to-end inference latency while maintaining superior task performance, demonstrating its practicality for resource-constrained LVLM deployment.
Abstract:Visual token pruning is a promising approach to reducing the inference cost of large vision-language models (LVLMs), yet aggressive token reduction often causes substantial performance degradation. We identify three key factors behind this degradation: text-guided selection bias, information loss from discarded tokens, and positional distortion caused by sequence compaction. Based on these observations, we propose \textbf{VPRune}, a training-free pre-LLM pruning framework consisting of visual-only diversity selection, similarity-guided token recycling, and position-preserving restoration. Experiments on FastVLM-1.5B across multiple vision-language benchmarks demonstrate that VPRune achieves a favorable accuracy--compression trade-off, with particularly pronounced advantages under aggressive compression. Furthermore, evaluations on edge-device show that VPRune effectively reduces end-to-end inference latency while maintaining superior task performance, demonstrating its practicality for resource-constrained LVLM deployment.
Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
arXiv:2609.24485 [cs.CV]
(or arXiv:2609.24485v2 [cs.CV] for this version)
https://doi.org/10.48550/arXiv.2609.24485 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 216 cuvinte
Subiecte conexe
[AINews] not much happened today
Latent Space · 10 sept. 2026
Clustering and Token Denoising for Faster and More Robust VLMs
arXiv · 21 aug. 2026
AnchorPrune: Relevance-Anchored Contextual Expansion for Visual Token Pruning
arXiv · 24 iul. 2026
VLA-ACL: Action-Consistent Visual Token Pruning for Efficient Vision-Language-Action Models
arXiv · 7 oct. 2026