mlsanityData to 7 Oct 2026

CONDUIT: A Unified Residual-Stream Restoration Framework for KV Cache Reuse in Vision-Language Models

arXiv 2609.05821 ↗First posted 5 Sep 2026cs.CL cs.CVcited by 0

On the map: Multimodal LLMs / vision-language models → Visual token pruning and efficiency

also in 1 more
Abstract (arXiv)

Vision-language models (VLMs) often answer new questions about recurring visual content, where reusing the key-value (KV) cache can avoid re-encoding expensive visual prefixes. Exact-prefix reuse, however, fails when the same visual content appears under a changed prefix. Selective recomputation can recover quality under a small visual-token budget, but only when the right stale tokens are refreshed. Raw-attention selection can waste budget on high-attention tokens with small value-norm proxy scores and on query-irrelevant images. To address these failure modes, we propose CONDUIT, a training-free refresh policy that unifies single- and multi-image reuse as residual-stream restoration. Building on norm-weighted attention, CONDUIT ranks cached visual tokens using cached-key query attention and an accessible pre-output cached-value-norm proxy, then applies empirical image-level relevance amplification before one global selection. With one image, the coefficient is one and the rule reduces to intra-image token selection. The method preserves model architecture and weights, adding only a single query-conditioned scoring pass at inference. At a 10% refresh budget, CONDUIT achieves 97.0-99.5% of the corresponding full-prefill five-dataset average across three VLM backbones and leads budgeted methods on average; on the MMLongBench-Doc latency subset, it uses 13.5% of full-prefill FLOPs and achieves a 2.99x time-to-first-token speedup.

Key facts. No citing papers in the map yet (data through 7 Oct 2026). Placed in Visual token pruning and efficiency, one of 16 directions in Multimodal LLMs / vision-language models. Builds on 39 of its 57 references that are in the map. First posted on arXiv 5 Sep 2026.

Builds on

39 of 57 references.

Closest ancestors · 5

Same direction, newest first

  1. 2026ViCA: Efficient Multimodal LLMs with Vision-Only Cross-Attention7 Feb
  2. 2025CASA: Cross-Attention over Self-Attention for Efficient Vision-Language Fusion22 Dec
  3. 2025MEDA: Dynamic KV Cache Allocation for Efficient Multimodal Long-Context Inference24 Feb
  4. 2025MPIC: Position-Independent Multimodal Context Caching System for Efficient MLLM Serving4 Feb
  5. 2024LOOK-M: Look-Once Optimization in KV Cache for Efficient Multimodal Long-Context Inference26 Jun

Other references · 34

  1. 2020An Image is Worth 16x16 Words: Transformers for Image Recognition at Scalecited by 29,915
  2. 2023Efficient Memory Management for Large Language Model Serving with PagedAttentioncited by 5,767
  3. 2024LLaVA-OneVision: Easy Visual Task Transfercited by 2,669
  4. 2023Efficient Streaming Language Models with Attention Sinkscited by 2,115
  5. 2023MMBench: Is Your Multi-modal Model an All-around Player?cited by 2,105
  6. 2025InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Modelscited by 1,769
  7. 2020Transformer Feed-Forward Layers Are Key-Value Memoriescited by 1,307
  8. 2024OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environmentscited by 1,134
  9. 2023SGLang: Efficient Execution of Structured Language Model Programscited by 1,107
  10. 2021DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsificationcited by 849
  11. 2024SnapKV: LLM Knows What You are Looking for Before Generationcited by 848
  12. 2020Quantifying Attention Flow in Transformerscited by 759
22 more
  1. 2024KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cachecited by 604
  2. 2022Transformer Feed-Forward Layers Build Predictions by Promoting Concepts in the Vocabulary Spacecited by 522
  3. 2023Scissorhands: Exploiting the Persistence of Importance Hypothesis for LLM KV Cache Compression at Test Timecited by 428
  4. 2024PyramidKV: Dynamic KV Cache Compression based on Pyramidal Information Funnelingcited by 420
  5. 2016Visual Dialogcited by 377
  6. 2024DuoAttention: Efficient Long-Context LLM Inference with Retrieval and Streaming Headscited by 283
  7. 2024Ada-KV: Optimizing KV Cache Eviction by Adaptive Budget Allocation for Efficient LLM Inferencecited by 230
  8. 2024MANTIS: Interleaved Multi-Image Instruction Tuningcited by 212
  9. 2024CacheBlend: Fast Large Language Model Serving for RAG with Cached Knowledge Fusioncited by 211
  10. 2023Can Pre-trained Vision and Language Models Answer Visual Information-Seeking Questions?cited by 203
  11. 2023SlideVQA: A Dataset for Document Visual Question Answering on Multiple Imagescited by 175
  12. 2024MMLongBench-Doc: Benchmarking Long-context Document Understanding with Visualizationscited by 174
  13. 2023Quantizable Transformers: Removing Outliers by Helping Attention Heads Do Nothingcited by 147
  14. 2024LongDocURL: a Comprehensive Multimodal Long Document Benchmark Integrating Understanding, Reasoning, and Locatingcited by 70
  15. 2025VDocRAG: Retrieval-Augmented Generation over Visually-Rich Documentscited by 57
  16. 2024MADTP: Multimodal Alignment-Guided Dynamic Token Pruning for Accelerating Vision-Language Transformercited by 53
  17. 2024Benchmarking Multi-Image Understanding in Vision and Language Models: Perception, Knowledge, Reasoning, and Multi-Hop Reasoningcited by 50
  18. 2025Attention Sinks and Compression Valleys in LLMs are Two Sides of the Same Coincited by 47
  19. 2025FastKV: Decoupling of Context Reduction and KV Cache Compression for Prefill-Decoding Accelerationcited by 13
  20. 2024Cross-Self KV Cache Pruning for Efficient Vision-Language Inferencecited by 11
  21. 2025KeepKV: Achieving Periodic Lossless KV Cache Compression for Efficient LLM Inferencecited by 8
  22. 2026QCFuse: Query-Aware Cache Fusion via Compressed View for Efficient RAG Servingcited by 2

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026Reroute, Don't Remove: Recoverable Visual Token Routing for Vision-Language Models
  2. 2025AirCache: Activating Inter-modal Relevancy KV Cache Compression for Efficient Large Vision-Language Model Inference
  3. 2026ReToken: One Token to Improve Vision-Language Models for Visual Retrieval
  4. 2026VLMaxxing through FrameMogging Training-Free Anti-Recomputation for Video Vision-Language Models
  5. 2025VScan: Rethinking Visual Token Reduction for Efficient Large Vision-Language Models
  6. 2026Adaptive Visual Token Reduction for Accelerated Image Understanding
  7. 2026Kamera: Unified Position-Invariant Multimodal KV Cache for Training-Free Reuse
  8. 2024VL-Cache: Sparsity and Modality-Aware KV Cache Compression for Vision-Language Model Inference Acceleration