Lune

CVPR Findings2026

InternVL-X: Advancing and Accelerating InternVL Series with Efficient Visual Token Compression

Dongchen Lu, Zilu Zhang, Leping Huang, Yuyao Sun, Jianliang Zeng, Mao Shu, Huo Cao

Abstract

Most multimodal large language models (MLLMs) treat visual tokens as "a sequence of text", integrating them with text tokens into a large language model (LLM). However, a great quantity of visual tokens significantly increases computational costs and latency. In this paper, we propose InternVL-X, which improves upon InternVL model in both performance and efficiency with three visual token compression methods. First, we introduce PVTC, a novel visionlanguage projector that aggregates spatially adjacent visual tokens into local queries and adopts the CLS token as global queries. By performing cross-attention between visual features and both the local and global queries, PVTC enables more effective extraction and transformation of visual embeddings. Second, the LVTC module compresses visual tokens at shallow layer of the LLM and decompresses them at deep layer, significantly enhancing computational efficiency and token utilization. It also employs multi-layer visual information fusion to improve performance. Furthermore, we propose RVTC, an efficient high-resolution slicing method that adjusts the number of visual tokens based on image pixels or edge length matching, greatly reducing computational resources consumption. By utilizing 18% or fewer visual tokens, InternVL-X achieves state-of-the-art performance on 11 public MLLM benchmarks, and improves the average metric by 2.65% across 12 tasks. Additionally, our approach boosts training efficiency by 43%, while maintaining superior model performance.

Ask Lune

Know this paper in minutes, not an afternoon.

Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.

7 top-tier papers have cited it since. Know what they found before your reviewers do.

Questions to start from

Lune can draw on

the full paper · the top-tier papers that cite it · related work

Get the answer

Free to start. No credit card required.

Lineage

  • An Efficient Token Compression Framework for Visual Object Tracking cites An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
  • An Efficient Token Compression Framework for Visual Object Tracking cites BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models
  • R1-ShareVL: Incentivizing Reasoning Capabilities of Multimodal Large Language Models via Share-GRPO cites Visual Instruction Tuning
  • Visual Instruction Tuning cites BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models
  • Visual Instruction Tuning cites Learning Transferable Visual Models From Natural Language Supervision
  • EventFlash: Towards Efficient MLLMs for Event-Based Vision cites Visual Instruction Tuning
  • EventFlash: Towards Efficient MLLMs for Event-Based Vision cites Learning Transferable Visual Models From Natural Language Supervision
  • Learning Transferable Visual Models From Natural Language Supervision cites An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
  • DeepEyes: Incentivizing "Thinking with Images" via Reinforcement Learning cites Visual Instruction Tuning
  • DeepEyes: Incentivizing "Thinking with Images" via Reinforcement Learning cites BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models
Builds onCited byLeft to right by year; the more cited, the higher.

Cited by top-tier papers7

  1. DeepEyes: Incentivizing "Thinking with Images" via Reinforcement LearningZiwei Zheng, Michael Yang, Jack Hong, Chenxiao Zhao et al.ICLR 2026 · 321 citations
  2. EventFlash: Towards Efficient MLLMs for Event-Based VisionShaoyu Liu, Jianing Li, Guanghui Zhao, Yunjian Zhang et al.ICLR 2026 · 5 citations
  3. AstraNav-Memory: Contexts Compression for Long MemoryJunjun Hu, Xinda Xue, Botao Ren, Minghua Luo et al.CVPR 2026 · 5 citations
  4. R1-ShareVL: Incentivizing Reasoning Capabilities of Multimodal Large Language Models via Share-GRPOHuanjin Yao, Qixiang Yin, Jingyi Zhang, Min Yang et al.NeurIPS 2025 · 3 citations
  5. An Efficient Token Compression Framework for Visual Object TrackingWeijing Wu, Qihua Liang, Bineng Zhong, Haiying Xia et al.CVPR 2026 · 1 citation

Builds on29

  1. Learning Transferable Visual Models From Natural Language SupervisionAlec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh et al.ICML 2021 · 47,906 citations
  2. An Image is Worth 16x16 Words: Transformers for Image Recognition at ScaleAlexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn et al.ICLR 2021 · 21,477 citations
  3. LoRA: Low-Rank Adaptation of Large Language ModelsEdward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu et al.ICLR 2022 · 18,833 citations
  4. Visual Instruction TuningHaotian Liu, Chunyuan Li, Qingyang Wu, Yong Jae LeeNeurIPS 2023 · 11,349 citations
  5. BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language ModelsJunnan Li, Dongxu Li, Silvio Savarese, Steven C. H. HoiICML 2023 · 7,873 citations

Related papers

  1. Efficient Large Multi-modal Models via Visual Context CompressionJieneng Chen, Luoxin Ye, Ju He, Zhaoyang Wang et al.NeurIPS 2024 · 49 citations
  2. Accelerating Multimodal Large Language Models by Searching Optimal Vision Token ReductionShiyu Zhao, Zhenting Wang, Felix Juefei-Xu, Xide Xia et al.CVPR 2025
  3. iLLaVA: An Image is Worth Fewer Than 1/3 Input Tokens in Large Multimodal ModelsLianyu Hu, Liqing Gao, Fanhua Shang, Liang Wan et al.ICLR 2026 · 9 citations
  4. Accelerating Multimodal Large Language Models via Dynamic Visual-Token Exit and the Empirical FindingsQiong Wu, Wenhao Lin, Yiyi Zhou, Weihao Ye et al.NeurIPS 2025 · 16 citations
  5. FlashSloth : Lightning Multimodal Large Language Models via Embedded Visual CompressionBo Tong, Bokai Lai, Yiyi Zhou, Gen Luo et al.CVPR 2025
Dusk over the sea between two cliffs drawn in fine vertical lines