Pith. sign in

REVIEW 4 major objections 4 minor 31 references

Modularized Dynamic-Granularity Video LLM for Multi-Event Long Video Understanding

T0 review · 4 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Iteratively grounding question-relevant segments, then encoding them finely and the rest coarsely, yields state-of-the-art results on multi-event long-video QA.

desk verdict A promising but under-verified modular recipe for long-video QA; reported gains may be DPO artifacts rather than better grounding. read the letter →

arxiv 2607.15778 v1 pith:E5AVXN6T submitted 2026-07-17 cs.CV cs.AI

Wei Feng , Xin Wang , Yu-Wei Zhan , Yuwei Zhou , Wenwu Zhu This is my paper
classification cs.CVcs.AI
keywords longvideounderstandinglargelanguagemodelstemporalgroundingtokengranularitymulti-eventQAdirectpreferenceoptimizationquestionansweringbenchmarkconstruction
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

MoD-VLLM's central claim is that the limiting factor in long-video QA is not the total token budget but how tokens are distributed across time. The paper proposes a closed loop: first, a grounding module instructs the video LLM to label segments as question-relevant (positive) or irrelevant (negative); then a modularized scheduler re-encodes the video so that positive segments receive fine-grained tokens and negative segments coarse-grained tokens; the resulting representation replaces the original uniform tokens and the loop repeats. To keep localization errors from snowballing, the first iteration also generates a counterfactual token sequence with reversed polarity and merges the two grounding outputs. Training uses DPO, where candidate grounding policies are preferred if they lead to lower cross-entropy on the answer when the video is encoded under that policy. The paper reports gains over strong 7B-8B baselines on VideoMME, Lvbench, and MLVU, and builds MEventBench (1,200 pairs, videos 15-100 min) on which the method leads, especially on multi-event counting.

What carries the argument

The modularized dynamic-granularity encoder scheduler (Algorithm 1) is the object that carries the argument: it takes the positive/negative segment split from grounding and, subject to a token budget L_max, assigns each segment an encoding module chosen from a set of fine-tuned projection layers with different per-frame token rates (coarse 36, medium 64, fine 169 tokens/frame in the implementation). The scheduler's granularity levels are functions of the positive fraction ρ (Eq. 6), and it first spends budget on positive segments, downgrading negative segments if over budget. The counterfactual sequence Z'' (Eq. 12) and the DPO objective (Eq. 13) support the loop by adding error correction a

What would settle it

Run MoD-VLLM on MEventBench with the first-iteration counterfactual sequence Z'' removed and with the grounding module replaced by a random segment split (matched for positive fraction). If accuracy stays similar, or if the grounding output's agreement with human-annotated relevant segments is poor, the claimed mechanism of dynamic granularity iteration is not the source of the gains. More directly, compute grounding precision/recall against human annotations on a sample of MEventBench; large grounding errors with high final accuracy would contradict the paper's explanation.

Watch

Extended reading notes

Core claim

The central discovery, as the authors state it, is that representing a long video with a single uniform token granularity—or with a one-shot keyframe selection—is inferior to an iterative process in which the model first identifies segments relevant to the question, then re-encodes those segments with more tokens (fine-grained) and the rest with fewer (coarse-grained), and repeats. The claimed mechanism is a closed loop of 'Positive-Negative Video Segments Grounding' and 'Modularized Dynamic-Granularity Reflection,' with a counterfactual granularity sequence used in the first iteration to limit error propagation from missed segments. They further claim that jointly optimizing grounding polic

Load-bearing premise

The paper's gains rest on the grounding step's JSON segment list being accurate enough to direct fine-grained tokens at the segments the question truly depends on; if grounding is frequently wrong, the reported improvements might come from the DPO answer-loss signal or dataset composition rather than from the dynamic-granularity allocation.

Editorial extensions

If this is right

  • Existing token-budget-limited video LLMs can improve long-video QA without scaling up model size, by reallocating tokens according to question relevance.
  • The largest reported gain on MEventBench is on multi-event counting (69.4% vs 64.2% for the strongest baseline), indicating that iterative grounding is especially beneficial when events are sparse and dispersed.
  • Jointly using visual tokens and question-frame similarity scores for grounding is necessary for the reported performance; ablating either modality drops accuracy, with similarity alone the weakest.
  • The framework outperforms both token-compression methods (LongVU) and two-stage keyframe-selection methods (AKS, VideoTree), suggesting that the iterative loop mitigates the error propagation these methods suffer.
  • DPO with answer cross-entropy as an implicit reward learns grounding policies without segment-level annotations, using only the final answer's loss to score candidate segment splits.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A directly testable corollary: the gain of dynamic granularity should increase with video length and with the sparsity of question-relevant events; the reported VideoMME Long gap (66.9 vs 62.6) is consistent, but a controlled sweep over video length would confirm it.
  • The counterfactual validation (Z'') is a credible safeguard but is not ablated; removing it in a controlled run would isolate how much of the improvement comes from error correction versus from the baseline grounding-DPO pair.
  • The scheduling formulas assume grounding quality is already decent; if grounding is noisy, a missed segment is coarsely encoded and may stay missed. An extension would be to add uncertainty-aware grounding or to re-encode segments that receive conflicting positive/negative labels across iterations.
  • The same loop could apply to other sequential modalities (e.g., audio tracks or text transcripts in a video) by swapping the encoder modules; the modular scheduler architecture does not depend on vision-specific details.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes MoD-VLLM, an iterative framework for multi-event long-video understanding. It introduces a Positive-Negative Video Segments Grounding module that uses a Video LLM to output question-relevant segments in JSON, and a Modularized Dynamic-Granularity Reflection module that encodes positive segments with fine-grained tokens and negative segments with coarse-grained tokens under a token budget. Grounding and reflection alternate over three iterations, with a counterfactual reversed-granularity sequence used in the first iteration to mitigate error propagation. A DPO-based reinforcement learning stage optimizes grounding-policy generation and dynamic-granularity representation jointly. The paper also contributes MEventBench, a 1200-pair multi-event benchmark filtered from existing long-video benchmarks. Experiments on VideoMME, Lvbench, MLVU, and MEventBench report consistent improvements over baselines.

Significance. If the claims hold, the work would be a useful contribution to long-video understanding: the grounding-plus-dynamic-granularity loop is a plausible way to allocate limited visual tokens to question-relevant segments, and MEventBench could serve as a focused evaluation set for multi-event reasoning. The paper also ships an algorithmic design (Algorithm 1 and Algorithm 2) with enough detail to be reproduced. However, the current evidence is not sufficient to attribute the reported gains to the core mechanism: the key components are not ablated, grounding accuracy is never measured, and the DPO reward is self-referential. The benchmark is not released, and no statistical significance is reported. These gaps must be addressed before the claims can be accepted.

major comments (4)
  1. [§IV.B, Table III] The only mechanism-level ablation varies input modalities (V/S). There is no ablation that removes the iterative refinement loop, the counterfactual validation of Eq. (12), the DPO stage of Algorithm 2, or the dynamic-granularity scheduler (replacing it with uniform fine-grained encoding). Consequently, the central claim that dynamic-granularity iteration improves multi-event QA is not directly supported; the gains in Table I could come from the SFT/DPO training on VideoITG or from benchmark composition rather than from correct positive/negative segment selection.
  2. [§III.C, Eqs. (13)/(21), Algorithm 2] The DPO preferences are computed from the cross-entropy loss of the very model being trained, on policies that the model generates itself (including random and counterfactual policies). This is self-referential: a policy that reduces answer loss may do so through token-budget or prompt-format artifacts rather than by selecting the true event segments. The paper never reports grounding precision/recall/IoU for Eq. (5) on any labeled dataset (e.g., VideoITG, which the authors already use for training). The authors should provide direct grounding-accuracy metrics and, ideally, show that answer loss is positively correlated with grounding quality, or use independent reward labels.
  3. [§III.B, Eqs. (6)-(9), Algorithm 1] The modularized scheduler is presented as solving the optimization in Eq. (9), but Algorithm 1 is a greedy heuristic with no convergence or optimality guarantee. Several hand-set parameters — α=1.5, preference weights w_p and w_n, token rates (36/64/169), and L_max — are never varied. Since the dynamic-granularity schedule is the load-bearing mechanism, the paper should include a sensitivity analysis of these choices and justify that the greedy procedure is an adequate approximation to Eq. (9).
  4. [§IV (MEventBench) and Supplement B] MEventBench is claimed as a contribution, but it is a filtered subset of existing benchmarks (VideoMME, Longvideobench, InfiniBench, CG-Bench), is not released, and has no inter-annotator agreement or quality analysis beyond a manual inspection step. Moreover, the main benchmark comparisons in Table I report single runs without error bars or significance tests; several gains are small (e.g., VideoMME Overall 73.2 vs. 71.6 for Qwen2.5-VL). The authors should release MEventBench, provide statistical significance, and report per-seed variance.
minor comments (4)
  1. [Table I] The VideoTree row is malformed: '* - - 54.2 - - - 55.4' is missing entries and the footnote about GPT-4 is unclear. Also, the paper claims 'significantly outperforms' but no confidence intervals are given anywhere.
  2. [Eq. (13) vs. Eq. (21)] The DPO objective in the main text conditions on (q, v), while the supplement version conditions only on q. Please align the notation and define the exact prompt conditioning used in Algorithm 2.
  3. [Throughout] Typographical issues: 'LLaV A-Video' should likely be 'LLaVA-Video'; 'continuous' in Supplement Eq. (19) should be 'continuous'; Figure 3 and some equations have missing spaces (e.g., 'V idLLM' in Algorithm 2).
  4. [Supplement B] The data distribution figure is referenced but not included in this version; the relationship between the 'several thousand candidates' and the final '1200 pairs' is unclear, and the duration ranges in Table I are not aligned with the described filtering thresholds.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: external benchmarks carry the empirical claim; the self-referential DPO reward is a training-bootstrap concern, not a definitional reduction.

full rationale

The paper's derivation chain is: positive-negative grounding (Eq. 5) → dynamic-granularity encoding (Eqs. 6–10) → counterfactual validation (Eq. 12) → DPO optimization (Eqs. 13/21) → evaluation on external benchmarks (VideoMME, Lvbench, MLVU, MEventBench). The only potentially self-referential element is the DPO preference construction: 'we apply DPO [21], [22] using the cross-entropy scores as implicit preferences,' where the 'positive' policy is selected by the lowest cross-entropy of the model's own answer. This is a self-training-style learning signal and could in principle be gamed by token-budget or format artifacts, but it is not a case of fitting a parameter and then 'predicting' the same fitted quantity. The paper's headline empirical claims are tested on held-out, externally defined benchmarks, not on the training objective itself, so the central claim does not reduce by construction to its inputs. The self-citations in the paper ([12], [13], [22]) are prior-work or inspiration citations; the DPO method itself is anchored in the standard external reference [21], and no uniqueness theorem or unverified ansatz is imported as load-bearing evidence. The absence of a direct grounding-precision/recall measurement and the lack of an ablation for the counterfactual validation (Eq. 12) are real evidence-quality limitations, but they concern mechanism attribution and correctness risk, not circularity of the derivation.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The central claim rests on several training/scheduling assumptions plus benchmark construction choices. The most fragile are grounding accuracy and the CE-loss-as-reward bootstrap. No code or data release is provided, so the empirical support is not independently checkable.

free parameters (7)
  • Granularity schedule exponent alpha = 1.5
    Eqs. 6/15; adjustable in [1.5,2.0], set to 1.5 without sensitivity analysis; controls how strongly positive/negative granularity levels separate.
  • Token rates of encoding modules = 36/64/169 tokens per frame
    Section IV/supplement; hand-set module capacities; the entire scheduling trade-off is defined by these rates.
  • Preference weights w_p, w_n in scheduler optimization = unspecified
    Eqs. 9/18; the module-selection objective depends on these weights, but their values are not reported.
  • DPO scaling beta = unspecified
    Eqs. 13/21; the strength of the preference optimization is not reported.
  • Number of dynamic-granularity iterations = 3
    Section IV: 'We run three dynamic granularity iterations per video'; no ablation of this number is given.
  • Number of sampled grounding policies per sample (n) = unspecified
    Algorithm 2 requires n but no value or detailed sampling strategy is given; affects the DPO candidate pool.
  • Token budget L_max and upgrade threshold = not stated numerically
    Eqs. 8/17; the scheduler constraint is central, but L_max and the 'far below' limit for upgrading granularity are not quantified.
assumptions (6)
  • domain assumption A token sequence that encodes question-relevant frames finely and irrelevant frames coarsely is more helpful for video LLM understanding.
    Supplement A, after Eq. 20: 'we can assert that a dynamic visual tokens sequence correctly represents... would be more helpful'. This is the core premise of the granularity allocation.
  • domain assumption The video LLM can produce reliable multi-segment JSON grounding outputs from coarse visual tokens plus CLIP similarity.
    Eq. 5 and Section III-A; no grounding accuracy metric is reported.
  • domain assumption The counterfactual reversed-granularity sequence reduces error propagation.
    Section III-C: 'To mitigate error propagation... we introduce a counterfactual validation'; asserted without ablation.
  • ad hoc to paper Granularity modules with more tokens/frame and higher sampling rate contain richer visual information.
    Section III-B: 'if p>n, then encoding module phi_p would contain richer visual information than phi_n' — defined by construction, not shown to be semantically richer.
  • domain assumption MEventBench manual filtering guarantees that questions indeed require multiple dispersed segments and that annotations are correct.
    Supplement B: manual inspection; no inter-annotator agreement and no public release.
  • domain assumption Cross-entropy loss on the generated answer is a valid implicit reward for grounding-policy quality.
    Section III-C/Supplement A: the preferred policy is the one with smallest CE; this assumes answer loss tracks grounding quality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Modularized Dynamic-Granularity Video LLM for Multi-Event Long Video Understanding." pith.science (2026). https://pith.science/paper/E5AVXN6T

@misc{pith2026260715778,
  author       = {Pith},
  title        = {Pith review of: Modularized Dynamic-Granularity Video LLM for Multi-Event Long Video Understanding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E5AVXN6T}},
  note         = {Machine review of arXiv:2607.15778}
}
read the original abstract

Video Large Language Models (Video LLMs) have made significant advancements in various video understanding tasks. However, long-video scenarios remain challenging due to the tension between limited visual token budgets and the need to capture multiple key events. Existing approaches typically process long videos in two stages, i.e., i) select keyframes and ii) perform detailed perception, which exhibit limitations: they lack a modular mechanism for adaptive capacity allocation and self-correction, resulting in unreliable modeling. To tackle these challenges, we propose MoD-VLLM, a novel Modularized Dynamic-Granularity Video LLM framework for multi-event long video understanding, which unifies temporal grounding and semantic understanding iteratively and self-reflectively. Specifically, we propose a Positive-Negative Video Segments Grounding module and a Modularized Dynamic-Granularity Reflection module, which form a closed loop to progressively localize the question-related video segments. The grounding module instructs a Video LLM to distinguish relevant from irrelevant video segments based on the video question. The reflection module employs a modularized scheduler that dynamically selects fine-grained encoding for relevant positive segments to capture detailed perception and coarse-grained encoding for negative segments to maintain global context. We further propose a dynamic-granularity reinforcement learning strategy, allowing MoD-VLLM to learn optimal grounding policies and dynamic granularity visual representation jointly. Moreover, we propose MEventBench, a challenging Multi-Event Long Video Benchmark for complex long video reasoning. Extensive experiments on several long video understanding benchmarks and our MEventBench demonstrate that MoD-VLLM significantly outperforms state-of-the-art baselines.

Figures

Figures reproduced from arXiv: 2607.15778 by the authors.

Figure 1
Figure 1. Conceptual comparison of different video LLM paradigms. Token [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Our MoD-VLLM framework for multi-event long video understanding. The positive-negative video segments grounding module instructs the video [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. A qualitative example of multi-event long video understanding. The [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Data distribution of MEventBench. Counting tasks. These data require models to identify and enumerate multiple instances of specific events or objects distributed sparsely throughout long videos. The key challenge lies in accurately detecting and tallying occurrences t…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 4 linked inside Pith

  1. [1]

    Visual instruction tuning,

    H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tuning,” in NeurIPS, vol. 36, 2023, pp. 34 892–34 916

    2023

  2. [2]

    Vtimellm: Empower llm to grasp video moments,

    B. Huang, X. Wang, H. Chen, Z. Song, and W. Zhu, “Vtimellm: Empower llm to grasp video moments,” inCVPR, 2024, pp. 14 271– 14 280

    2024

  3. [3]

    Moviechat: From dense token to sparse memory for long video understanding,

    E. Song, W. Chai, G. Wang, Y . Zhang, H. Zhou, F. Wu, H. Chi, X. Guo, T. Ye, Y . Zhanget al., “Moviechat: From dense token to sparse memory for long video understanding,” inCVPR, 2024, pp. 18 221–18 232

    2024

  4. [4]

    Longvu: Spatiotemporal adaptive compression for long video-language understanding,

    X. Shen, Y . Xiong, C. Zhao, L. Wu, J. Chen, C. Zhu, Z. Liu, F. Xiao, B. Varadarajan, F. Bordeset al., “Longvu: Spatiotemporal adaptive compression for long video-language understanding,” inICML, 2025

    2025

  5. [5]

    Adaptive keyframe sampling for long video understanding,

    X. Tang, J. Qiu, L. Xie, Y . Tian, J. Jiao, and Q. Ye, “Adaptive keyframe sampling for long video understanding,” inCVPR, 2025, pp. 29 118– 29 128

    2025

  6. [6]

    Videotree: Adaptive tree-based video representation for llm reasoning on long videos,

    Z. Wang, S. Yu, E. Stengel-Eskin, J. Yoon, F. Cheng, G. Bertasius, and M. Bansal, “Videotree: Adaptive tree-based video representation for llm reasoning on long videos,” inCVPR, 2025, pp. 3272–3283

    2025

  7. [7]

    Llama: Open and efficient foundation language models,

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozi `ere, N. Goyal, E. Hambro, F. Azharet al., “Llama: Open and efficient foundation language models,”arXiv preprint arXiv:2302.13971, 2023

    arXiv 2023

  8. [8]

    Multi-modal generative ai: Multi-modal llms, diffusions and the unification,

    X. Wang, Y . Zhou, B. Huang, H. Chen, and W. Zhu, “Multi-modal generative ai: Multi-modal llms, diffusions and the unification,”IEEE Transactions on Circuits and Systems for Video Technology, 2025

    2025

Show all 31 references
  1. [9]

    Video-llama: An instruction-tuned audio- visual language model for video understanding,

    H. Zhang, X. Li, and L. Bing, “Video-llama: An instruction-tuned audio- visual language model for video understanding,” inEMNLP, 2023, pp. 543–553

    2023

  2. [10]

    Fuyu-8b: A multimodal architecture for ai agents,

    R. Bavishi, E. Elsen, C. Hawthorne, M. Nye, A. Odena, A. Somani, and S. Tas ¸ırlar, “Fuyu-8b: A multimodal architecture for ai agents,” 2023

    2023

  3. [11]

    X. Wang, X. Lan, and W. Zhu,Video Grounding and Its Generalization: From ID and Task-specific Models to OOD and Large Foundation Models. Springer, 2025

    2025

  4. [12]

    Multi-sentence video grounding for long video generation,

    W. Feng, X. Wang, H. Chen, Z. Zhang, and W. Zhu, “Multi-sentence video grounding for long video generation,” inICME. IEEE, 2025, pp. 1–6

    2025

  5. [13]

    Modularagent: A task-aware modular framework for joint optimization of multimodal large language models and world models,

    Y .-W. Zhan, X. Wang, P. Mao, T. Feng, R. Wang, and W. Zhu, “Modularagent: A task-aware modular framework for joint optimization of multimodal large language models and world models,” inCVPR, 2026

    2026

  6. [14]

    Video-rag: Visually-aligned retrieval-augmented long video comprehension,

    Y . Luo, X. Zheng, X. Yang, G. Li, H. Lin, J. Huang, J. Ji, F. Chao, J. Luo, and R. Ji, “Video-rag: Visually-aligned retrieval-augmented long video comprehension,”arXiv preprint arXiv:2411.13093, 2024

    2024

  7. [15]

    Video-xl: Extra-long vision language model for hour-scale video understanding,

    Y . Shu, Z. Liu, P. Zhang, M. Qin, J. Zhou, Z. Liang, T. Huang, and B. Zhao, “Video-xl: Extra-long vision language model for hour-scale video understanding,” inCVPR, 2025, pp. 26 160–26 169

    2025

  8. [16]

    Video-xl-2: Towards very long-video understanding through task-aware kv sparsification,

    M. Qin, X. Liu, Z. Liang, Y . Shu, H. Yuan, J. Zhou, S. Xiao, B. Zhao, and Z. Liu, “Video-xl-2: Towards very long-video understanding through task-aware kv sparsification,”arXiv preprint arXiv:2506.19225, 2025

    2025 arXiv

  9. [17]

    Vita-1.5: Towards gpt-4o level real-time vision and speech interaction,

    C. Fu, H. Lin, X. Wang, Y .-F. Zhang, Y . Shen, X. Liu, H. Cao, Z. Long, H. Gao, K. Liet al., “Vita-1.5: Towards gpt-4o level real-time vision and speech interaction,”arXiv preprint arXiv:2501.01957, 2025

    2025 arXiv

  10. [18]

    Qwen2. 5-vl technical report,

    S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tanget al., “Qwen2. 5-vl technical report,”arXiv preprint arXiv:2502.13923, 2025

    2025 arXiv

  11. [19]

    Llava-video: Video instruction tuning with synthetic data,

    Y . Zhang, J. Wu, W. Li, B. Li, Z. MA, Z. Liu, and C. Li, “Llava-video: Video instruction tuning with synthetic data,”Transactions on Machine Learning Research

  12. [20]

    Deep reinforcement learning from human preferences,

    P. F. Christiano, J. Leike, T. Brown, M. Martic, S. Legg, and D. Amodei, “Deep reinforcement learning from human preferences,”Advances in neural information processing systems, vol. 30, 2017

    2017

  13. [21]

    Direct preference optimization: Your language model is secretly a reward model,

    R. Rafailov, A. Sharma, E. Mitchell, C. D. Manning, S. Ermon, and C. Finn, “Direct preference optimization: Your language model is secretly a reward model,”NeurIPS, vol. 36, pp. 53 728–53 741, 2023

    2023

  14. [22]

    Modularized self-reflected video reasoner for multimodal llm with application to video question answering,

    Z. Song, X. Wang, Z. Qian, H. Chen, L. Huang, H. Xue, and W. Zhu, “Modularized self-reflected video reasoner for multimodal llm with application to video question answering,” inICML, 2025

    2025

  15. [23]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” inICML. PMLR, 2021, pp. 8748–8763

    2021

  16. [24]

    Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis,

    C. Fu, Y . Dai, Y . Luo, L. Li, S. Ren, R. Zhang, Z. Wang, C. Zhou, Y . Shen, M. Zhanget al., “Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis,” inCVPR, 2025, pp. 24 108–24 118

    2025

  17. [25]

    Lvbench: An extreme long video understanding benchmark,

    W. Wang, Z. He, W. Hong, Y . Cheng, X. Zhang, J. Qi, M. Ding, X. Gu, S. Huang, B. Xuet al., “Lvbench: An extreme long video understanding benchmark,” inICCV, 2025, pp. 22 958–22 967

    2025

  18. [26]

    Mlvu: Benchmarking multi-task long video understanding,

    J. Zhou, Y . Shu, B. Zhao, B. Wu, Z. Liang, S. Xiao, M. Qin, X. Yang, Y . Xiong, B. Zhanget al., “Mlvu: Benchmarking multi-task long video understanding,” inCVPR, 2025, pp. 13 691–13 701

    2025

  19. [27]

    Longvideobench: A benchmark for long-context interleaved video-language understanding,

    H. Wu, D. Li, B. Chen, and J. Li, “Longvideobench: A benchmark for long-context interleaved video-language understanding,”NeurIPS, vol. 37, pp. 28 828–28 857, 2024

    2024

  20. [28]

    Infinibench: A comprehensive benchmark for large multimodal models in very long video understanding,

    K. Ataallah, C. Gou, E. Abdelrahman, K. Pahwa, J. Ding, and M. Elhoseiny, “Infinibench: A comprehensive benchmark for large multimodal models in very long video understanding,”arXiv preprint arXiv:2406.19875, 2024

    2024

  21. [29]

    Cg-bench: Clue-grounded question answering benchmark for long video understanding,

    G. Chen, Y . Liu, Y . Huang, B. Pei, J. Xu, Y . He, T. Lu, Y . Wang, and L. Wang, “Cg-bench: Clue-grounded question answering benchmark for long video understanding,” inICLR, 2025

    2025

  22. [30]

    Videoitg: Multimodal video understanding with instructed temporal grounding,

    S. Wang, G. Chen, D.-a. Huang, Z. Li, M. Li, G. Li, J. M. Alvarez, L. Zhang, and Z. Yu, “Videoitg: Multimodal video understanding with instructed temporal grounding,”arXiv preprint arXiv:2507.13353, 2025. SUPPLEMENT A. Training details The overall training of our framework is ...

    2025

  23. [31]

    ordering,

    Algorithm details:Modularized encoding scheduler. Compared to the original uni-granularity representation, we hope to represent the positive video segment as fine-grained as possible with a higher ratio of sampled frames, but not exceeding the token limitationsL max1 of the vi...

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.