F^3A: Training-Free Visual Token Pruning for Multimodal Language Models

A recent study published on arXiv (2605.16359) presents F^3A, a router designed for visual token pruning in vision-language models without requiring training. This technique tackles the issue of determining the necessary number of visual tokens as multimodal models expand, while also managing their allocation within a set budget. Unlike current training-free pruning strategies that rely on one-time proxies such as decoder attention or visual similarity, F^3A conceptualizes visual token pruning as a task-conditioned evidence search, particularly useful for high compression and various model sizes. It functions prior to the language model processing image tokens, creating lightweight, question-conditioned cues and matching them to visual-grid tokens using frozen sparse sensing heads, thereby managing a fixed vision token budget through coarse evidence localization. The authors assert that this method is superior for scenarios requiring significant compression.