PoRe: Position-Reweighted Visual Token Pruning for Vision Language Models
This work addresses a specific bottleneck in making VLMs more efficient, offering an incremental but practical plug-and-play solution for researchers and practitioners.
The paper tackles the problem of recency bias in visual token pruning for Vision-Language Models, where inflated attention scores for bottom image tokens lead to suboptimal pruning; their position-reweighting method improves pruning performance with minimal computational overhead, as shown in experiments on LVLMs.
Vision-Language Models (VLMs) typically process a significantly larger number of visual tokens compared to text tokens due to the inherent redundancy in visual signals. Visual token pruning is a promising direction to reduce the computational cost of VLMs by eliminating redundant visual tokens. The text-visual attention score is a widely adopted criterion for visual token pruning as it reflects the relevance of visual tokens to the text input. However, many sequence models exhibit a recency bias, where tokens appearing later in the sequence exert a disproportionately large influence on the model's output. In VLMs, this bias manifests as inflated attention scores for tokens corresponding to the lower regions of the image, leading to suboptimal pruning that disproportionately retains tokens from the image bottom. In this paper, we present an extremely simple yet effective approach to alleviate the recency bias in visual token pruning. We propose a straightforward reweighting mechanism that adjusts the attention scores of visual tokens according to their spatial positions in the image. Our method, termed Position-reweighted Visual Token Pruning, is a plug-and-play solution that can be seamlessly incorporated into existing visual token pruning frameworks without any changes to the model architecture or extra training. Extensive experiments on LVLMs demonstrate that our method improves the performance of visual token pruning with minimal computational overhead.