mlsanityData to 14 Oct 2026

TwinPrune: Density-Aware Two-Phase Token Pruning for Vision-Language Models

arXiv 2610.17530 ↗First posted 13 Oct 2026cs.CVcited by 0

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

also in 1 more
Abstract (arXiv)

We propose a unified spatial-density signal to score token importance for both inference phases: bipartite merge for prefill and KV eviction for decode, yielding measurable gains in each stage. To isolate the two phases, we split benchmarks by scoring logic. Prefill probes (MMBench, POPE, ScienceQA) evaluate only the first decoded token; their accuracy reflects prefill logits argmax before any decode KV eviction. Full-pipeline probes (TextVQA, DocVQA, MM-Vet) assess the complete generated sequence and capture both phases. We identify a critical evaluation pitfall: decode pruning claims validated solely on prefill probes measure prefill quality, not decode performance. Experiments show discarding all visual KV during decode barely affects prefill-probe accuracy but degrades TextVQA and MM-Vet drastically; this insensitivity stems from the scoring rule, not compression intensity. Evaluated on seven model variants (InternVL-3.5, Qwen3-VL, LLaVA-1.5) and three full-pipeline benchmarks, our method achieves a three-way Pareto frontier across compute, memory and accuracy. Density-based prefill compression maintains strong accuracy, scales linearly with visual token count, and delivers 100x faster token selection at high resolution. For decode, density top-k KV eviction consistently outperforms strong baselines H2O, SnapKV and StreamingLLM on TextVQA and MM-Vet across three retention ratios. At 0.125 keep ratio on TextVQA, it surpasses H2O by 9.6 points; InternVL-3.5 models (2B to 38B) maintain a 7 to 10 point advantage. Compression scales favorably with model size: every measured 8B to 38B model transition reduces accuracy loss instead of worsening it. We suggest MM-Vet or DocVQA as mandatory full-pipeline validation for any decode-phase claims. Our code is available at https://github.com/Kai-Liu001/TwinPrune

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in Visual token pruning and efficiency, one of 16 directions in Multimodal LLMs / vision-language models. Builds on 28 of its 33 references that are in the map. First posted on arXiv 13 Oct 2026.

Builds on

28 of 33 references.

Closest ancestors · 10

Same direction, newest first

  1. 2025Similarity-Aware Token Pruning: Your VLM but Faster14 Mar
  2. 2025DivPrune: Diversity-based Visual Token Pruning for Large Multimodal Models4 Mar
  3. 2025Stop Looking for Important Tokens in Multimodal Language Models: Duplication Matters More17 Feb
  4. 2024VisionZip: Longer is Better but Not Necessary in Vision Language Models5 Dec
  5. 2024Fit and Prune: Fast and Training-free Visual Token Pruning for Multi-modal Large Language Models16 Sep
  6. 2024Efficient Inference of Vision Instruction-Following Models with Elastic Cache25 Jul
  7. 2024TokenPacker: Efficient Visual Projector for Multimodal LLM2 Jul
  8. 2024LOOK-M: Look-Once Optimization in KV Cache for Efficient Multimodal Long-Context Inference26 Jun
  9. 2024LLaVA-PruMerge: Adaptive Token Reduction for Efficient Large Multimodal Models22 Mar
  10. 2024An Image is Worth 1/2 Tokens After Layer 2: Plug-and-Play Inference Acceleration for Large Vision-Language Models11 Mar

Other references · 18

  1. 2023Improved Baselines with Visual Instruction Tuningcited by 4,418
  2. 2025Qwen3-VL Technical Reportcited by 2,793
  3. 2023Efficient Streaming Language Models with Attention Sinkscited by 2,186
  4. 2023MMBench: Is Your Multi-modal Model an All-around Player?cited by 2,141
  5. 2022Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answeringcited by 1,919
  6. 2023Evaluating Object Hallucination in Large Vision-Language Modelscited by 1,729
  7. 2019Towards VQA Models That Can Readcited by 1,718
  8. 2025InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiencycited by 1,510
  9. 2023MM-Vet: Evaluating Large Multimodal Models for Integrated Capabilitiescited by 1,239
  10. 2022Token Merging: Your ViT But Fastercited by 1,021
  11. 2024SnapKV: LLM Knows What You are Looking for Before Generationcited by 880
  12. 2023H2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Modelscited by 832
6 more
  1. 2024KIVI: A Tuning-Free Asymmetric 2bit Quantization for KV Cachecited by 637
  2. 2024Quest: Query-Aware Sparsity for Efficient Long-Context LLM Inferencecited by 501
  3. 2023Model Tells You What to Discard: Adaptive KV Cache Compression for LLMscited by 474
  4. 2014Lazier Than Lazy Greedycited by 99
  5. 2017Practical Coreset Constructions for Machine Learningcited by 78
  6. 2025R-KV: Redundancy-aware KV Cache Compression for Reasoning Modelscited by 54

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026AsymVLM: Asymmetric Token Pruning for Efficient Vision-Language Model Inference
  2. 2025A Glimpse to Compress: Dynamic Visual Token Pruning for Large Vision-Language Models
  3. 2026Towards Joint Quantization and Token Pruning of Vision-Language Models
  4. 2025TopV: Compatible Token Pruning with Inference Time Optimization for Fast and Low-Memory Multimodal Vision Language Model
  5. 2026From Patching to Pruning Visual Computation in Vision Language Models
  6. 2024Feather the Throttle: Revisiting Visual Token Pruning for Vision-Language Model Acceleration
  7. 2025Balanced Token Pruning: Accelerating Vision Language Models Beyond Local Optimization
  8. 2026VPRune: Efficient Training-free Pre-LLM Visual Token Pruning