Chain-of-Thought Compression Should Not Be Blind: V-Skip for Efficient Multimodal Reasoning via Dual-Path Anchoring
This addresses efficiency issues for users of multimodal AI systems by improving reasoning speed without sacrificing accuracy, though it is an incremental advance over existing compression methods.
The paper tackles the problem of high latency in Chain-of-Thought reasoning for Multimodal Large Language Models by introducing V-Skip, a method that prevents visual hallucinations during token compression, achieving a 2.9x speedup with minimal accuracy loss and outperforming baselines by over 30% on DocVQA.
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.