Vision token pruning improves by picking best method per image
Beyond One-Size-Fits-All: Sample-Adaptive Strategy Routing for Vision Token Pruning in MLLMs
Computer Vision and Pattern RecognitionArtificial Intelligence
Summary
Processing images with large AI models is expensive because they use many visual pieces called tokens. The authors found that no single pruning method works best for every image—some methods work better on certain pictures. They created Vip-Router, a small system that picks the best pruning method for each image using simple clues, saving effort without losing accuracy. This approach works well across different models and images without changing the main AI itself.
What this means in practice
- •For ai system engineers: Reduce visual processing cost by dynamically choosing the best token pruning method per image to improve AI inference efficiency and accuracy.
- •For mobile app developers: Improve image understanding in apps by integrating adaptive pruning to lower compute requirements on diverse images without losing output quality.
Authors
Haiji Liang, Pengfei Zhou, Zhenglin Wan, Wei Wang, Yang You, Wangbo Zhao
Abstract
Multimodal large language models (MLLMs) process hundreds or thousands of visual tokens per image, incurring prohibitive inference costs. While existing vision token pruning methods mitigate this overhead, they implicitly assume that a single fixed pruning strategy can be applied uniformly across all inputs. Our analysis further reveals that ranking pruning methods by average benchmark accuracy conceals substantial sample-wise complementarity: although the average-best strategy excels overall, alternative strategies prove superior on a significant fraction of individual samples. To harness this diversity, we propose VIP-Router, a lightweight VIsion Pruning Router that adaptively selects the pruning strategy predicted to be best suited to each input at a specified pruning level. Conditioned on low-cost visual and textual features, VIP-Router identifies the most suitable candidate strategy while retaining full-token inference as an option when pruning is predicted to be unfavorable. Evaluated on a curated suite of pruning-sensitive visual perception benchmarks, VTC-Bench Group A, VIP-Router consistently outperforms the best fixed strategy baseline across all reduction ratios, achieving a 26.9% relative improvement in average accuracy, and a 22.0% relative increase in average utility after accounting for realized token cost. Crucially, VIP-Router operates in a plug-and-play manner without modifying underlying pruning algorithms or model weights, introducing trainable parameters equivalent to merely 0.017\% of the backbone. Furthermore, VIP-Router proves effective across various MLLM backbones and yields consistent gains on unseen benchmarks, highlighting the potential of sample adaptive routing for visual token pruning.