Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

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

About

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.

Jianwei Zhang, Chaoning Zhang, Sihan Cao, Wang Liu, Pengcheng Zheng, Jiaxin Huang, Caiyan Qin, Yalan Ye, Wei Dong, Yang Yang• 2026

Related benchmarks

TaskDatasetResultRank
Visual Question AnsweringVizWiz
Accuracy99.06
1525
Object Hallucination EvaluationPOPE--
1455
Visual Question AnsweringGQA
GQA Score60.2
85
Multimodal BenchmarkingMMB
Average Performance63.61
40
Multimodal EvaluationMME
Absolute Score1.79e+3
20
Visual Question AnsweringVQA v2
Absolute Score77.54
20
Visual Question AnsweringPOPE v2
Accuracy99.06
15
Multimodal Model EvaluationMME
MME Score1.79e+3
5
Showing 8 of 8 rows

Other info

Follow for update