Chain-of-Thought Compression Should Not Be Blind: V-Skip for Efficient Multimodal Reasoning via Dual-Path Anchoring
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2026
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866911505524457472 |
|---|---|
| author | Zhang, Dongxu Sun, Yiding Tan, Cheng Yan, Wenbiao Yang, Ning Zhu, Jihua Zhang, Haijun |
| author_facet | Zhang, Dongxu Sun, Yiding Tan, Cheng Yan, Wenbiao Yang, Ning Zhu, Jihua Zhang, Haijun |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2601_13879 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Chain-of-Thought Compression Should Not Be Blind: V-Skip for Efficient Multimodal Reasoning via Dual-Path Anchoring Zhang, Dongxu Sun, Yiding Tan, Cheng Yan, Wenbiao Yang, Ning Zhu, Jihua Zhang, Haijun Multimedia Computation and Language Computer Vision and Pattern Recognition 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. |
| title | Chain-of-Thought Compression Should Not Be Blind: V-Skip for Efficient Multimodal Reasoning via Dual-Path Anchoring |
| topic | Multimedia Computation and Language Computer Vision and Pattern Recognition |
| url | https://arxiv.org/abs/2601.13879 |