RCP: Representation Consistency Pruner for Mitigating Distribution Shift in Large Vision-Language Models

TL;DR

RCP is a novel pruning framework for large vision-language models that reduces visual tokens and computational costs while maintaining performance through a delayed repair mechanism.

Contribution

It introduces a cross-attention based cumulative pruning method with a delayed repair adapter, enabling efficient token reduction without fine-tuning the entire model.

Findings

Removes up to 88.9% of visual tokens

Reduces FLOPs by up to 85.7%

Maintains performance with marginal accuracy drop

Abstract

Large Vision-Language Models (LVLMs) suffer from prohibitive inference costs due to the massive number of visual tokens processed by the language decoder. Existing pruning methods often lead to significant performance degradation because the irreversible removal of visual tokens causes a distribution shift in the hidden states that deviates from the pre-trained full-token regime. To address this, we propose Representation Consistency Pruner, which we refer to as RCP, as a novel framework that integrates cumulative visual token pruning with a delayed repair mechanism. Specifically, we introduce a cross-attention pruner that leverages the intrinsic attention of the LLM as a baseline to predict cumulative masks, ensuring consistent and monotonic token reduction across layers. To compensate for the resulting information loss, we design a delayed repair adapter denoted as DRA, which caches the essence of pruned tokens and applies FiLM-based modulation specifically to the answer generation tokens. We employ a repair loss to match the first and second-order statistics of the pruned representations with a full-token teacher. RCP is highly efficient because it trains only lightweight plug-in modules while allowing for physical token discarding at inference. Extensive experiments on LVLM benchmarks demonstrate that RCP removes up to 88.9\% of visual tokens and reduces FLOPs by up to 85.7\% with only a marginal average accuracy drop, and outperforms prior methods that avoid fine-tuning the original model on several widely used benchmarks.