mlsanityData to 7 Oct 2026

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

arXiv 2601.13879 ↗First posted 20 Jan 2026cs.MM cs.CL cs.CVcited by 21

On the map: LLM reasoning → Multimodal reasoning

also in 2 more
Abstract (arXiv)

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× speedup with negligible accuracy loss. Specifically, it preserves fine-grained visual details, outperforming other baselines over 30% on the DocVQA.

Key facts. Cited by 21 papers in the map (data through 7 Oct 2026). Placed in Multimodal reasoning, one of 17 directions in LLM reasoning. Builds on 17 of its 34 references that are in the map. First posted on arXiv 20 Jan 2026.

Builds on

17 of 34 references.

Closest ancestors · 2

Same direction, newest first

  1. 2025Multimodal Chain-of-Thought Reasoning: A Comprehensive Survey16 Mar
  2. 2023Multimodal Chain-of-Thought Reasoning in Language Models2 Feb

Other references · 15

  1. 2018Representation Learning with Contrastive Predictive Codingcited by 7,662
  2. 2024Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolutioncited by 4,151
  3. 2023Evaluating Object Hallucination in Large Vision-Language Modelscited by 1,696
  4. 2022In-context Learning and Induction Headscited by 930
  5. 2016Deep Variational Information Bottleneckcited by 915
  6. 2019Fast Transformer Decoding: One Write-Head is All You Needcited by 700
  7. 2018Object Hallucination in Image Captioningcited by 661
  8. 2024LLaVA-NeXT-Interleave: Tackling Multi-image, Video, and 3D in Large Multimodal Modelscited by 571
  9. 2024LLMLingua-2: Data Distillation for Efficient and Faithful Task-Agnostic Prompt Compressioncited by 343
  10. 2023LLMLingua: Compressing Prompts for Accelerated Inference of Large Language Modelscited by 279
  11. 2025TokenSkip: Controllable Chain-of-Thought Compression in LLMscited by 246
  12. 2024Learning to Compress Prompt in Natural Language Formatscited by 54
3 more
  1. 2025Token-Importance Guided Direct Preference Optimizationcited by 14
  2. 2025Not All Errors Are Created Equal: ASCoT Addresses Late-Stage Fragility in Efficient LLM Reasoningcited by 11
  3. 2025Breaking the MoE LLM Trilemma: Dynamic Expert Clustering with Structured Compressioncited by 3

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 21.

21 papers.

Sort:

By area:

  1. 2026PointCoT: A Multi-modal Benchmark for Explicit 3D Geometric Reasoningcited by 19
  2. 2026Not All Queries Need Deep Thought: CoFiCot for Adaptive Coarse-to-fine Stateful Refinementcited by 13
  3. 2026PointRFT: Explicit Reinforcement Fine-tuning for Point Cloud Few-shot Learningcited by 10
  4. 2026CMHANet: A Cross-Modal Hybrid Attention Network for Point Cloud Registrationcited by 9
  5. 2026IGASA: Integrated Geometry-Aware and Skip-Attention Modules for Enhanced Point Cloud Registrationcited by 5
  6. 2026Pansharpening for Thin-Cloud Contaminated Remote Sensing Images: A Unified Framework and Benchmark Datasetcited by 5
  7. 2026Robust Regularized Policy Iteration under Transition Uncertaintycited by 2
  8. 2026CFMS: A Coarse-to-Fine Multimodal Synthesis Framework for Enhanced Tabular Reasoningcited by 2
  9. 2026FADE: Mitigating Hallucinations by Reducing Language-Prior Dominance in Large Vision-Language Modelscited by 2
  10. 2026SeeMe: Mitigating Hallucinations in Large Vision-Language Models through Effective Visual Token Engineeringcited by 2
  11. 2026Parameter-Efficient Fine-Tuning for Spiking Point Cloud Modelscited by 2
  12. 2026Can LLM-Generated Text Empower Surgical Vision-Language Pre-training?cited by 1
9 more
  1. 2026SF-AMS: Strategic Forgetting for Structured Memory in LLM Agentcited by 1
  2. 2026Bridging Efficiency and Transparency: Explainable CoT Compression in Multimodal Large Reasoning Modelscited by 0
  3. 2026Controllable Singing Style Conversion with Boundary-Aware Information Bottleneckcited by 0
  4. 2026Sparsity-Aware Voxel Attention and Foreground Modulation for 3D Semantic Scene Completioncited by 0
  5. 2026SPARK: Susceptibility-Guided Profiling and Steering of Latent Reasoning States in Large Language Modelscited by 0
  6. 2026Better Starts, Better Ends: Bootstrapped Iterative Self-Reasoning Distillation for Compressed Reasoningcited by 0
  7. 2026Buried in Textual Debt: Context Pruning with Visual Evidence Preservation for MLLM Agentscited by 0
  8. 2026Motion-Saliency Complementary Masked Modeling for Point Cloud Video Understandingcited by 0
  9. 2026DRT: Dense Reasoning Trace for Efficient and Grounded Multimodal Reasoningcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026VIG: Visual Information Gain as a Reward Signal for Multimodal Chain-of-Thought Compression
  2. 2026Look Less, Reason More: Block-wise Attention Skipping for Efficient Multimodal LLMs
  3. 2026Reflection Anchors for Propagation-Aware Visual Retention in Long-Chain Multimodal Reasoning
  4. 2026Learning Modal-Mixed Chain-of-Thought Reasoning with Latent Embeddings
  5. 2026VisionPulse: Dynamic Visual Sparsity for Efficient Multimodal Reasoning
  6. 2025Mitigating Visual Forgetting via Take-along Visual Conditioning for Multi-modal Long CoT Reasoning
  7. 2026Thinking with Images as Continuous Actions: Numerical Visual Chain-of-Thought
  8. 2026Attend, Transform, or Silence: Operator-Level Visual Skipping for Efficient Multimodal LLM Inference