Chain-of-Thought Compression Should Not Be Blind: V-Skip for Efficient Multimodal Reasoning via Dual-Path Anchoring

TL;DR

This paper introduces V-Skip, a dual-path gating method for efficient multimodal reasoning that balances token compression with visual information preservation, significantly speeding up processing while maintaining accuracy.

Contribution

V-Skip reformulates token pruning as a Visual-Anchored Information Bottleneck problem, effectively preventing visual information loss during compression in multimodal models.

Findings

V-Skip achieves a 2.9x speedup with negligible accuracy loss.

It outperforms baselines by over 30% on DocVQA.

Preserves fine-grained visual details in multimodal reasoning.

Abstract

While Chain-of-Thought (CoT) reasoning significantly enhances the performance of Multimodal Large Language Models (MLLMs), its autoregressive nature incurs prohibitive latency constraints. Current efforts to mitigate this via token compression often fail by blindly applying text-centric metrics to multimodal contexts. We identify a critical failure mode termed Visual Amnesia, where linguistically redundant tokens are erroneously pruned, leading to hallucinations. To address this, we introduce V-Skip that reformulates token pruning as a Visual-Anchored Information Bottleneck (VA-IB) optimization problem. V-Skip employs a dual-path gating mechanism that weighs token importance through both linguistic surprisal and cross-modal attention flow, effectively rescuing visually salient anchors. Extensive experiments on Qwen2-VL and Llama-3.2 families demonstrate that V-Skip achieves a $2.9\times$ speedup with negligible accuracy loss. Specifically, it preserves fine-grained visual details, outperforming other baselines over 30\% on the DocVQA.