mlsanityData to 14 Oct 2026

Seen, Said, or Forgotten? A Causal Audit of Visual KV Memory Across Dialog Turns

arXiv 2607.25467 ↗First posted 28 Jul 2026cs.CV cs.AIcited by 1

Not on the map. Why?

Abstract (arXiv)

Stateful multimodal assistants encode an image once but may answer questions about it many turns later. Attention-guided visual-KV eviction assumes that evidence irrelevant now will remain dispensable, although future questions are unknown. We ask when a visual fact is actually safe to forget and introduce the Causal Visual Memory Audit (CVMA), a paired single-prefill framework that tests what later answers lose when a visual region, the whole image, or prior assistant text becomes unavailable. On VisDial and ConvBench, current attention can rank future-useful regions worse than random even though a diagnostic marginal-utility control shows substantial selection headroom. Aggregate scores hide this failure when later turns do not need vision; controlled and stock-generated histories reveal a second escape route, in which assistant-text KV replaces image KV for facts already stated but not reliably for unstated facts. In the tested stacks, safe forgetting is supported by low future visual dependence or fact-specific verbalization---not by low current attention.

Key facts. Cited by 1 paper in the map (data through 14 Oct 2026). Builds on 21 of its 25 references that are in the map. First posted on arXiv 28 Jul 2026.

Builds on

21 of 25 references.

References · 21

  1. 2023Efficient Streaming Language Models with Attention Sinkscited by 2,186
  2. 2024SnapKV: LLM Knows What You are Looking for Before Generationcited by 880
  3. 2024KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cachecited by 637
  4. 2016Visual Dialogcited by 377
  5. 2024Boosting Multimodal Large Language Models with Visual Tokens Withdrawal for Rapid Inferencecited by 139
  6. 2024No Token Left Behind: Reliable KV Cache Compression via Importance-Aware Mixed Precision Quantizationcited by 101
  7. 2024LOOK-M: Look-Once Optimization in KV Cache for Efficient Multimodal Long-Context Inferencecited by 95
  8. 2024NACL: A General and Effective KV Cache Eviction Framework for LLMs at Inference Timecited by 43
  9. 2024VL-Cache: Sparsity and Modality-Aware KV Cache Compression for Vision-Language Model Inference Accelerationcited by 43
  10. 2025Memory-efficient Streaming VideoLLMs for Real-time Procedural Video Understandingcited by 19
  11. 2025SparseVILA: Decoupling Visual Sparsity for Efficient VLM Inferencecited by 15
  12. 2024PrefixKV: Adaptive Prefix KV Cache is What Vision Instruction-Following Models Need for Efficient Generationcited by 13
9 more
  1. 2026Make Each Token Count: Towards Improving Long-Context Performance with KV Cache Evictioncited by 11
  2. 2025FlowKV: Enhancing Multi-Turn Conversational Coherence in LLMs via Isolated Key-Value Cache Managementcited by 6
  3. 2026Predicting Future Utility: Global Combinatorial Optimization for Task-Agnostic KV Cache Evictioncited by 6
  4. 2025CalibQuant: 1-Bit KV Cache Quantization for Multimodal LLMscited by 3
  5. 2025Revisiting Multimodal KV Cache Compression: A Frequency-Domain-Guided Outlier-KV-Aware Approachcited by 3
  6. 2026SONIC: Segmented Optimized Nexus for Information Compression in Key-Value Cachingcited by 3
  7. 2026RetentiveKV: State-Space Memory for Uncertainty-Aware Multimodal KV Cache Evictioncited by 3
  8. 2026NestedKV: Nested Memory Routing for Long-Context KV Cache Compressioncited by 2
  9. 2026Last But Not Least: Boundary Attention CalibratiON for Multimodal KV Cache Compressioncited by 1

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 1.

1 paper.

  1. 2026PhoenixNest-Video: Evidence-Grounded Multimodal Agent Framework for Automated Video Interview Assessmentcited by 1