pith. sign in

arxiv: 2606.02161 · v1 · pith:NGFGFV7Dnew · submitted 2026-06-01 · 💻 cs.CV · cs.CL

InfoMerge: Information-aware Token Compression for Efficient Video Large Language Models

Xinxin Liu , Shiwei Gan , Xiao Liu , Yafeng Yin , Lei Xie , Sanglu Lu This is my paper

Pith reviewed 2026-06-28 15:03 UTC · model grok-4.3

classification 💻 cs.CV cs.CL
keywords video large language modelsvisual token compressiontemporal redundancy estimationcontent-aware budget allocationtraining-free efficiencyspectral entropytoken reduction
0
0 comments X

The pith

InfoMerge reduces visual tokens in video LLMs by 85 percent while retaining 98.8 percent of original performance through segment-level redundancy estimation and entropy-driven allocation.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper aims to show that current training-free token compression in Video-LLMs falls short because it relies on adjacent-frame similarity or segment length, both of which miss the uneven spread of information across real videos and get thrown off by noise. InfoMerge instead estimates redundancy at the segment level by tracking second-order changes in token fingerprints at fixed spatial positions and then assigns token budgets according to each segment's uniqueness measured by spectral entropy. If this approach holds, models can drop most visual tokens yet still answer questions about video content nearly as well as the uncompressed version, especially when compression ratios are high.

Core claim

InfoMerge is a training-free method that first computes Temporal Fingerprint Difference within each video segment to capture how tokens at the same spatial locations change over time in a second-order sense, then applies Content-Aware Budget Allocation that scores segments by their spectral-entropy richness and uniqueness so that more tokens are kept in informative parts and fewer in static or repeated regions. On LLaVA-OneVision-7B this yields 85 percent token reduction, 4.24 times faster prefill, and 98.8 percent retention of average benchmark performance, with the gains widening at aggressive compression rates across multiple backbones.

What carries the argument

Temporal Fingerprint Difference, a segment-level second-order temporal redundancy estimator that models similarity structure of same-position tokens across frames inside each segment, paired with Content-Aware Budget Allocation that distributes token quotas by spectral-entropy representational richness and segment uniqueness.

If this is right

  • Under 85 percent token reduction the method still keeps 98.8 percent average accuracy on standard video-understanding benchmarks.
  • Prefill-stage inference speeds up by a factor of 4.24 on the 7B LLaVA-OneVision backbone.
  • Advantages become larger when the allowed token budget is made even smaller.
  • The same compression pipeline transfers to other Video-LLM backbones without retraining.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same fingerprint-plus-entropy logic could be applied to compress audio tokens in multimodal models that handle both video and sound.
  • If the method scales, real-time video chat systems could process minute-long clips on consumer GPUs without dropping frames.
  • The approach suggests that information-aware allocation might also improve efficiency when compressing image tokens inside static vision-language models.

Load-bearing premise

That segment-level second-order fingerprint differences and spectral-entropy scores give a reliable, noise-resistant measure of information content that works across videos and model backbones without any extra tuning.

What would settle it

A controlled test set of videos where static background noise is added that changes pixel values but leaves semantic content unchanged; if InfoMerge then drops performance more sharply than length-based or frame-similarity baselines, the redundancy estimator is not robust.

Figures

Figures reproduced from arXiv: 2606.02161 by Lei Xie, Sanglu Lu, Shiwei Gan, Xiao Liu, Xinxin Liu, Yafeng Yin.

Figure 1
Figure 1. Figure 1: Qualitative analysis of token selection. (a) [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed InfoMerge method. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of the proposed TFD. For each [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Analysis of segment-wise low-rank struc￾tures in video token representations. The rapidly decaying singular-value spectra indicate strong redun￾dancy within the segment. 3.3 Content-Aware Budget Allocation Video information is often unevenly distributed across time: long segments may contain mostly static or repetitive content, while short segments can include key actions or scene changes. There￾fore, allo… view at source ↗
Figure 5
Figure 5. Figure 5: Ablation on the temperature parameter τ in dynamic budget allocation. Moderate temperature leads to the best trade-off between smooth and aggressive budget redistribution. Retain Ratio = 10%. Effect of TFD penalty λ on VideoMME 57.0 57.2 57.4 57.6 57.8 58.0 1 2 5 8 10 12 15 10⁹ TFD penalty λ VideoMME Acc (%) 57.3 57.1 57.3 57.4 57.7 57.7 57.7 57.7 ★ [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Ablation on the redundancy suppression strength λ on VideoMME. Stronger suppression reduces repeated preservation of static tokens and improves to￾ken utilization. Retain Ratio = 10%. controls the sensitivity of the sigmoid modulation function to the segment score. A smaller value makes the budget distribution smoother, while a larger value concentrates the token budget on a few high-scoring segments. As s… view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of token selection in the proposed InfoMerge.Segments with lower information content [PITH_FULL_IMAGE:figures/full_fig_p014_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Case study of evidence-aware budget allocation on VideoMME. The question asks what caused the sudden [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
read the original abstract

Video Large Language Models (Video-LLMs) achieve strong performance in video understanding, but their excessive visual tokens bring substantial computational overhead. Existing training-free compression methods improve inference efficiency by reducing visual tokens, yet they often rely on local adjacent-frame similarity for temporal redundancy estimation or allocate token budgets mainly according to segment length. Such designs are sensitive to frame-level noise and fail to capture the non-uniform information distribution of real-world videos. To address these challenges, we propose InfoMerge, a training-free visual token compression method that improves token utilization through robust redundancy estimation and content-aware budget allocation. Specifically, we propose the Temporal Fingerprint Difference: a segment-level second-order temporal redundancy estimation strategy, which models the temporal similarity structure of tokens at the same spatial positions within each segment. We further introduce Content-Aware Budget Allocation (CABA), which dynamically allocates segment-level token budgets based on segment uniqueness and spectral-entropy-based representational richness. By reducing repeated preservation of redundant static regions and allocating more tokens to informative segments, InfoMerge makes better use of the limited token budget while maintaining strong performance. Extensive experiments show that InfoMerge achieves strong efficiency--accuracy trade-offs across multiple benchmarks and backbones, with more pronounced advantages under aggressive compression. On LLaVA-OneVision-7B, InfoMerge retains 98.8\% of the original average performance while reducing 85\% of visual tokens and achieving a 4.24-fold speedup in the prefill stage.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces InfoMerge, a training-free visual token compression method for Video-LLMs. It proposes Temporal Fingerprint Difference, a segment-level second-order temporal redundancy estimator operating on per-position token sequences, and Content-Aware Budget Allocation (CABA) that distributes token budgets according to segment uniqueness combined with spectral-entropy richness. On LLaVA-OneVision-7B the method is reported to retain 98.8% of baseline average performance while discarding 85% of visual tokens and delivering a 4.24-fold prefill speedup.

Significance. If the proposed metrics prove robust, the work would offer a practical, training-free route to higher token efficiency in video understanding models, directly addressing the quadratic cost of long visual sequences. The emphasis on non-uniform information distribution and the absence of any learned parameters are positive features.

major comments (2)
  1. [Abstract and §3] Abstract and §3 (method description): the central performance claim (98.8% retention at 85% reduction) rests on the assumption that segment-level second-order Temporal Fingerprint Difference and spectral-entropy CABA reliably quantify information content; no ablation isolating sensitivity to camera motion, lighting variation, or encoder-specific token statistics is presented, leaving the robustness assumption load-bearing and untested.
  2. [§4] §4 (experiments): the reported benchmarks are confined to standard video QA and captioning suites; cross-backbone or cross-domain tests that would verify whether the second-order fingerprint and entropy metrics generalize without model-specific calibration are absent, directly affecting the generalization statement.
minor comments (2)
  1. Table captions should explicitly state the number of runs and any variance reported for the 98.8% retention figure.
  2. Notation for the second-order difference operator should be introduced with a short equation in §3.1 rather than only in prose.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback highlighting the need for stronger validation of our metrics' robustness and generalization. We address each major comment point by point below.

read point-by-point responses

  1. Referee: [Abstract and §3] Abstract and §3 (method description): the central performance claim (98.8% retention at 85% reduction) rests on the assumption that segment-level second-order Temporal Fingerprint Difference and spectral-entropy CABA reliably quantify information content; no ablation isolating sensitivity to camera motion, lighting variation, or encoder-specific token statistics is presented, leaving the robustness assumption load-bearing and untested.

    Authors: We agree that dedicated ablations isolating sensitivity to camera motion, lighting variation, and encoder-specific token statistics would strengthen the robustness claims for Temporal Fingerprint Difference and CABA. While the second-order design of the fingerprint metric is intended to reduce sensitivity to first-order perturbations such as lighting changes by modeling per-position temporal similarity structures, this remains an assumption without direct testing. We will add a new ablation subsection in the revised §4 with controlled experiments on videos exhibiting varying motion speeds, lighting conditions, and across different visual encoders to directly address this concern. revision: yes

  2. Referee: [§4] §4 (experiments): the reported benchmarks are confined to standard video QA and captioning suites; cross-backbone or cross-domain tests that would verify whether the second-order fingerprint and entropy metrics generalize without model-specific calibration are absent, directly affecting the generalization statement.

    Authors: The manuscript already reports results across multiple backbones (as noted in the abstract and §4) to support applicability beyond a single model. However, we acknowledge that the current experiments do not include explicit cross-domain tests on non-standard video distributions or direct verification that the metrics require no model-specific calibration. We will expand the experimental section with additional cross-backbone and cross-domain results in the revision to better substantiate the generalization statement. revision: partial

Circularity Check

0 steps flagged

No circularity: new metrics and allocation are defined independently of results

full rationale

The paper defines Temporal Fingerprint Difference as a segment-level second-order difference on per-position token sequences and CABA as a dynamic allocation using uniqueness plus spectral entropy; neither quantity is fitted to the target performance metric nor derived from a self-citation chain. The 98.8% retention claim is an empirical outcome on benchmarks, not a quantity that reduces to the input definitions by construction. No load-bearing self-citation, ansatz smuggling, or renaming of known results appears in the provided text.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Review performed on abstract only; no explicit free parameters, invented entities, or additional axioms beyond the stated domain motivation are identifiable.

axioms (1)
  • domain assumption Real-world videos exhibit non-uniform information distribution across segments and are sensitive to frame-level noise in local similarity measures.
    Invoked to justify the shift from local adjacent-frame similarity to segment-level second-order estimation and content-aware allocation.

pith-pipeline@v0.9.1-grok · 5804 in / 1304 out tokens · 33065 ms · 2026-06-28T15:03:53.338261+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 10 canonical work pages · 8 internal anchors

  1. [3]

    Science China Information Sciences , volume=

    Videochat: Chat-centric video understanding , author=. Science China Information Sciences , volume=. 2025 , publisher=

    2025

  2. [7]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Visionzip: Longer is better but not necessary in vision language models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  3. [8]

    European Conference on Computer Vision , pages=

    An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models , author=. European Conference on Computer Vision , pages=. 2024 , organization=

    2024

  4. [9]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Dycoke: Dynamic compression of tokens for fast video large language models , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  5. [10]

    Advances in Neural Information Processing Systems , volume=

    Fastvid: Dynamic density pruning for fast video large language models , author=. Advances in Neural Information Processing Systems , volume=

  6. [13]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    D ^2 Pruner: Debiased Importance and Structural Diversity for MLLM Token Pruning , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  7. [14]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Mmg-vid: Maximizing marginal gains at segment-level and token-level for efficient video llms , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  8. [15]

    2007 15th European signal processing conference , pages=

    The effective rank: A measure of effective dimensionality , author=. 2007 15th European signal processing conference , pages=. 2007 , organization=

    2007

  9. [18]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Vila: On pre-training for visual language models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  10. [19]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Nvila: Efficient frontier visual language models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  11. [21]

    International Conference on Learning Representations , volume=

    Tempme: Video temporal token merging for efficient text-video retrieval , author=. International Conference on Learning Representations , volume=

  12. [22]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    Prunevid: Visual token pruning for efficient video large language models , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  13. [23]

    LLaVA-NeXT: Improved reasoning, OCR, and world knowledge , url=

    Liu, Haotian and Li, Chunyuan and Li, Yuheng and Li, Bo and Zhang, Yuanhan and Shen, Sheng and Lee, Yong Jae , month=. LLaVA-NeXT: Improved reasoning, OCR, and world knowledge , url=

  14. [24]

    Advances in neural information processing systems , volume=

    Attention is all you need , author=. Advances in neural information processing systems , volume=

  15. [25]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Mvbench: A comprehensive multi-modal video understanding benchmark , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  16. [26]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Mlvu: Benchmarking multi-task long video understanding , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  17. [27]

    Advances in Neural Information Processing Systems , volume=

    Longvideobench: A benchmark for long-context interleaved video-language understanding , author=. Advances in Neural Information Processing Systems , volume=

  18. [28]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  19. [29]

    Lmms-eval: Accelerating the development of large multimoal models , volume=

    Xinrun du, yuhao dong, haotian liu, yuanhan zhang, ge zhang, chunyuan li, and ziwei liu , author=. Lmms-eval: Accelerating the development of large multimoal models , volume=

  20. [30]

    Findings of the Association for Computational Linguistics: NAACL 2025 , pages=

    Lmms-eval: Reality check on the evaluation of large multimodal models , author=. Findings of the Association for Computational Linguistics: NAACL 2025 , pages=

    2025

  21. [31]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    On perfect clustering of high dimension, low sample size data , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2019 , publisher=

    2019

  22. [32]

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, and 1 others. 2025. Qwen3-vl technical report. arXiv preprint arXiv:2511.21631

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  23. [33]

    Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. 2022. Token merging: Your vit but faster. arXiv preprint arXiv:2210.09461

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  24. [34]

    Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. 2024. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. In European Conference on Computer Vision, pages 19--35. Springer

    2024

  25. [35]

    Zesen Cheng, Sicong Leng, Hang Zhang, Yifei Xin, Xin Li, Guanzheng Chen, Yongxin Zhu, Wenqi Zhang, Ziyang Luo, Deli Zhao, and 1 others. 2024. Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms. arXiv preprint arXiv:2406.07476

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  26. [36]

    Junhao Du, Jialong Xue, Anqi Li, Jincheng Dai, and Guo Lu. 2026. Unified spatiotemporal token compression for video-llms at ultra-low retention. arXiv preprint arXiv:2603.21957

    work page arXiv 2026

  27. [37]

    Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, and 1 others. 2025. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24108--24118

    2025

  28. [38]

    Xiaohu Huang, Hao Zhou, and Kai Han. 2025. Prunevid: Visual token pruning for efficient video large language models. In Findings of the Association for Computational Linguistics: ACL 2025, pages 19959--19973

    2025

  29. [39]

    Bo Li. 2024. Xinrun du, yuhao dong, haotian liu, yuanhan zhang, ge zhang, chunyuan li, and ziwei liu. Lmms-eval: Accelerating the development of large multimoal models, 6

    2024

  30. [40]

    Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and 1 others. 2024 a . Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  31. [41]

    Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. 2024 b . Llava-next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  32. [42]

    KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. 2025. Videochat: Chat-centric video understanding. Science China Information Sciences, 68(10):200102

    2025

  33. [43]

    Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. 2024. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 26689--26699

    2024

  34. [44]

    Zhijian Liu, Ligeng Zhu, Baifeng Shi, Zhuoyang Zhang, Yuming Lou, Shang Yang, Haocheng Xi, Shiyi Cao, Yuxian Gu, Dacheng Li, and 1 others. 2025. Nvila: Efficient frontier visual language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4122--4134

    2025

  35. [45]

    Junpeng Ma, Qizhe Zhang, Ming Lu, Zhibin Wang, Qiang Zhou, Jun Song, and Shanghang Zhang. 2026. Mmg-vid: Maximizing marginal gains at segment-level and token-level for efficient video llms. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 24253--24261

    2026

  36. [46]

    Soham Sarkar and Anil K Ghosh. 2019. On perfect clustering of high dimension, low sample size data. IEEE transactions on pattern analysis and machine intelligence, 42(9):2257--2272

    2019

  37. [47]

    Kele Shao, Keda Tao, Can Qin, Haoxuan You, Yang Sui, and Huan Wang. 2025. Holitom: Holistic token merging for fast video large language models. arXiv preprint arXiv:2505.21334

    work page arXiv 2025

  38. [48]

    Leqi Shen, Guoqiang Gong, Tao He, Yifeng Zhang, Pengzhang Liu, Sicheng Zhao, and 1 others. 2026. Fastvid: Dynamic density pruning for fast video large language models. Advances in Neural Information Processing Systems, 38:123553--123581

    2026

  39. [49]

    Leqi Shen, Tianxiang Hao, Tao He, Sicheng Zhao, Yifeng Zhang, Yongjun Bao, Guiguang Ding, and 1 others. 2025. Tempme: Video temporal token merging for efficient text-video retrieval. In International Conference on Learning Representations, volume 2025, pages 60839--60860

    2025

  40. [50]

    Keda Tao, Can Qin, Haoxuan You, Yang Sui, and Huan Wang. 2025. Dycoke: Dynamic compression of tokens for fast video large language models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 18992--19001

    2025

  41. [51]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30

    2017

  42. [52]

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, and 1 others. 2024. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  43. [53]

    Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Ye, Jie Shao, and 1 others. 2025. Internvl3. 5: Advancing open-source multimodal models in versatility, reasoning, and efficiency. arXiv preprint arXiv:2508.18265

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  44. [54]

    Haoning Wu, Dongxu Li, Bei Chen, and Junnan Li. 2024. Longvideobench: A benchmark for long-context interleaved video-language understanding. Advances in Neural Information Processing Systems, 37:28828--28857

    2024

  45. [55]

    Senqiao Yang, Yukang Chen, Zhuotao Tian, Chengyao Wang, Jingyao Li, Bei Yu, and Jiaya Jia. 2025. Visionzip: Longer is better but not necessary in vision language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19792--19802

    2025

  46. [56]

    Evelyn Zhang, Fufu Yu, Aoqi Wu, Zichen Wen, Ke Yan, Shouhong Ding, Biqing Qi, and Linfeng Zhang. 2026. D ^2 pruner: Debiased importance and structural diversity for mllm token pruning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 12412--12420

    2026

  47. [57]

    Kaichen Zhang, Bo Li, Peiyuan Zhang, Fanyi Pu, Joshua Adrian Cahyono, Kairui Hu, Shuai Liu, Yuanhan Zhang, Jingkang Yang, Chunyuan Li, and 1 others. 2025. Lmms-eval: Reality check on the evaluation of large multimodal models. In Findings of the Association for Computational Linguistics: NAACL 2025, pages 881--916

    2025

  48. [58]

    Yuanhan Zhang, Jinming Wu, Wei Li, Bo Li, Zejun Ma, Ziwei Liu, and Chunyuan Li. 2024. Llava-video: Video instruction tuning with synthetic data. arXiv preprint arXiv:2410.02713

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  49. [59]

    Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Zhengyang Liang, Shitao Xiao, Minghao Qin, Xi Yang, Yongping Xiong, Bo Zhang, and 1 others. 2025. Mlvu: Benchmarking multi-task long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13691--13701

    2025