Pith. sign in

REVIEW 3 major objections 6 minor 51 references

Cross-modal attention from a frozen 2B MLLM suffices to select query-relevant frames, improving long-video QA by up to 6.4 points over uniform sampling under a 32-frame budget.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 22:34 UTC pith:3MWKSA7I

load-bearing objection A solid training-free frame-selection paper with a genuinely new attention-scoring trick, a mostly clean controlled comparison, and a calibration presentation that needs fixing before the no-leakage claim holds up. the 3 major comments →

arxiv 2607.15689 v1 pith:3MWKSA7I submitted 2026-07-17 cs.CV

Efficient Frame Selection for Long Videos at Test Time with Attention-Based MLLM Selectors

Yilin Wang , Xiangxi Zheng , Dongxing Mao , Linjie Li , Zhengyuan Yang , Ping Yu , Rui Yan , Yuan Yao
show 1 more author
Alex Jinpeng Wang
This is my paper
classification cs.CV
keywords frame selectionlong-video understandingmultimodal large language modelscross-modal attentiontraining-free selectionbudget-aware allocationdynamic programmingvideo question answering
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to break the circular dependency in long-video question answering: picking the right frames seems to require understanding the video, but understanding the video requires the right frames. The authors claim that the internal cross-modal attention of a multimodal LLM, extracted at a mid-depth layer chosen on validation data, already encodes query-relevant frame evidence before any answer generation. They build DAFS, a training-free selector that converts that attention into frame-level relevance scores and chooses top-K frames for a stronger answerer. Under a 32-frame budget this improves accuracy over uniform sampling by up to 6.4 points on Video-MME and beats prior training-based selectors at matched budgets, while transferring across selector and answerer backbones without retraining. If right, expensive autoregressive selectors and trained keyframe models are unnecessary for this task.

Core claim

The central claim is that cross-modal attention at validation-selected extraction layers in MLLMs already provides query-relevant frame evidence without requiring autoregressive generation. Concretely, text-to-vision attention weights from a selected layer are aggregated over each candidate frame's visual tokens, then max-pooled over the question's most informative tokens, to yield a per-frame saliency score; ranking by this score selects frames that improve a fixed answerer. The paper shows this signal is strongest in intermediate layers before deep cross-modal fusion blurs frame-level details, that the scores are query-conditioned rather than mere visual similarity, and that a 2B parameter

What carries the argument

The central object is the selected-layer text-to-vision attention matrix of a small MLLM selector, mapped from visual-token indices to frames. Per-frame saliency is the maximum over important query tokens of the summed attention mass on that frame's visual tokens; important tokens are answer-option tokens for multiple-choice and LLM-extracted entity tokens for open QA. The second mechanism is a dynamic-programming policy that, under a hard visual-token budget, chooses how many candidate frames and how many visual tokens per frame to use for each duration bin, with monotonicity constraints; the selected frames are then fed to the answerer at original resolution. The DP allocation handles the

Load-bearing premise

The candidate pool is built by uniformly subsampling a fixed number of frames from a base 1 FPS extraction, so the selector can only choose among those instants; if the decisive evidence falls between two sampled instants, no attention score can ever bring it back.

What would settle it

Take a long video containing a single decisive event shorter than one second that occurs exactly halfway between two consecutive 1 FPS sample instants, with no visible trace in either neighboring sampled frame. If DAFS with a 32-frame budget cannot select the event frame and the answerer fails while a selector that can see the event succeeds, the sampling premise is the bottleneck. Alternatively, measure oracle recall of DAFS scores against ground-truth frames on a benchmark where the answer depends on a sub-second event: if recall at the sampled instants is at chance, the attention signal is

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A frozen small MLLM can serve as a query-conditioned frame relevance scorer, so frame selection no longer needs autoregressive decoding, reinforcement learning, or keyframe supervision.
  • The selected frames transfer across answerers of very different sizes and across task formats (multiple-choice, open-ended, preview-retrieve-answer tool use), indicating the evidence is answerer-independent.
  • The gains are largest under tight frame budgets (e.g., +14.4 points on MLVU N-QA at 8 frames), so attention-based scoring is most valuable when only a few frames can be kept.
  • The DP allocation of pool size versus per-frame tokens is a real contributor: static allocations lose roughly half a point to over two points, so duration-aware budget calibration matters.
  • The selection stage is much cheaper than a trained autoregressive selector (0.47 s versus 8.00 s per query), so the accuracy improvement does not come at the price of inference cost.

Where Pith is reading between the lines

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

  • The paper leaves implicit that its selector is limited to frames from a uniform 1 FPS subsample; a natural extension is to score temporal segments or denser candidate pools, since sub-second decisive events between sampled instants are unrecoverable by any attention score.
  • The extraction layer itself is a calibration choice made on disjoint grounding benchmarks; an automatic criterion for choosing that layer without a separate validation set would make the method fully self-contained.
  • The same selected-layer attention could plausibly serve other temporal-localization tasks such as temporal grounding or highlight detection, since the paper shows it concentrates on query-relevant visual evidence at the frame level.
  • The max-pooling over important tokens suppresses positional bias but depends on the prompt containing salient entities or answer options; questions without explicit entities may need an alternative definition of importance.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper addresses test-time frame selection for long-video MLLM question answering. It proposes DAFS, which uses text-to-vision attention at a selected layer of a small frozen MLLM as a query-conditioned relevance score for uniformly subsampled candidate frames. A max-pooling aggregation over query-important tokens converts attention into frame-level saliency, and a dynamic-programming procedure calibrates duration-dependent allocations of candidate-pool size and per-frame token budget under a fixed selector context. The selected top-K frames are then fed to a stronger answerer at original resolution. Under a 32-frame budget with a fixed Qwen2.5-VL-7B answerer, DAFS reports gains over uniform sampling of up to +6.4 on Video-MME, +5.5 on LongVideoBench, and +11.5 on MLVU, and outperforms several training-based selectors. Additional experiments show cross-selector and cross-answerer transfer and a preview-retrieve-answer tool-use setting.

Significance. If the central claim holds, the paper makes a useful contribution: it demonstrates that cross-modal attention from a 2B-parameter MLLM, without autoregressive decoding, can serve as a practical query-relevant frame evidence signal, and that budget allocation can be calibrated offline. The strengths are the controlled comparisons in Table 2 (same answerer, same frame budget), the use of external grounding sets for layer selection (Charades-STA/NExT-GQA, Figure 4), the cross-backbone generalization in Table 5, and the wall-clock cost analysis in Table 3. These design choices make the empirical comparison more trustworthy than many prior selector papers. The remaining risk is calibration transparency: the paper claims no evaluation-data leakage, but some tables appear to use evaluation benchmarks in calibration or ablations, and the DP allocation is itself a second calibration knob whose contribution to the reported gains is not isolated. If those issues are resolved, the paper would be a solid empirical contribution to long-video frame selection.

major comments (3)
  1. [§4.1, Table 7] The validation protocol states that no selector parameter is tuned on evaluation benchmarks and that the extraction layer ℓ⋆ is selected by a Charades-STA/NExT-GQA sweep (Table 1 and Table 7). Table 7, however, is an ablation on Video-MME that reports per-layer accuracy and highlights the adopted setting (InternVL3.5-2B L12). If Video-MME accuracy informed the choice of L12, the headline gains in Table 2 are inflated by test-set selection; if it did not, the table should be relabeled as post-hoc analysis and the text should not cite it as part of the selection sweep. Table 1 similarly includes a Video-MME column inside the layer-calibration table. Please state unambiguously which tables were used to select layers, and remove or reposition any eval-benchmark columns from the calibration description.
  2. [§3.4, Eq. (4), §4.1] The DP allocation policy is a second calibration knob beyond the extraction layer. Eq. (4) maximizes held-out QA accuracy of the fixed answerer on 1,000 calibration videos from CG-Bench, NExT-QA, LongVideo-Reason, and LLaVA-Video. The no-leak statement in §3.1 covers only the layer sweep, not the DP stage. Because the reported gains in Table 2 use the DP-tuned allocation, those gains may partly reflect allocation regularities in the calibration QA datasets rather than attention relevance alone. Please report (i) calibration-set accuracy of the DP policy, (ii) the gains of a simple duration-aware hand heuristic under the same budget, and (iii) stability of the chosen (F(d), P(d)) across calibration-set splits. Without this, the relative contribution of the DP stage to the headline improvement is unclear.
  3. [§3.3] The candidate pool is constructed by uniformly subsampling F(d) frames from a 1 FPS base extraction. The attention scores can only rank frames inside this pool; if the decisive evidence occurs between sampled instants, no attention signal can recover it. This is a boundary condition on the central claim. The paper should state this limitation explicitly and, ideally, provide a sensitivity study over base sampling rates or F(d) values to show how the reported gains depend on candidate density. The current text does not acknowledge this dependence.
minor comments (6)
  1. [§4.1] Eq. (4) uses Acc_i(F_i, P_i) but the definition is incomplete. Please specify how Acc_i is computed (number of videos per duration bin, metric, and how the average over K ∈ {16,32,64} is taken). Also clarify whether the same calibrated policy is used for all K rows in Table 2 or whether separate policies are derived for each budget.
  2. [§3.2] The construction of T_imp is underspecified. For multiple-choice tasks, are all option tokens used, or only distractors, or the correct answer? For open-ended QA, which LLM extracts entity-centric tokens, with what prompt, and is this extraction performed per query at test time? Please provide details and, if possible, an ablation of T_imp composition.
  3. [§3.4] The monotonicity constraints (6)-(7) rule out by assumption a policy that gives longer videos more per-frame tokens on a short but critical segment. Please justify these constraints beyond intuition, or show that they do not materially restrict the solution.
  4. [Figure 4] The left panel labels the x-axis 'Selector depth' while the right panel uses 'Layer Index'. Align the terminology and clarify which model each panel refers to.
  5. [Table 5] Please state explicitly which selector is used for the cross-model generalization experiments (presumably InternVL3.5-2B with the L12 extraction layer). The caption currently says 'the same selector' without naming it.
  6. [General] The term 'training-free' is used despite offline calibration on QA labels. Consider replacing it with 'no gradient updates' or 'inference-time calibrated' to avoid confusion about the use of validation labels.

Circularity Check

0 steps flagged

No significant circularity: attention scoring and DP calibration are validated on disjoint data; reported gains are measured, not constructed.

full rationale

The derivation chain is not circular. The load-bearing claim is that text-to-vision attention from a fixed small MLLM, aggregated over query-important tokens (Eqs. 1-2), ranks frames that improve a stronger answerer. The scoring rule is defined independently of the evaluation benchmarks; its usefulness is first checked against ground-truth temporal grounding (Charades-STA/NExT-GQA) and then against measured QA accuracy on disjoint benchmarks (Video-MME, LVB, MLVU) under matched budgets. The extraction layer is fixed by the Charades-STA/NExT-GQA sweep; even though Table 1 also displays Video-MME numbers, the text explicitly states the layer is chosen from the disjoint grounding sweep, so no per-layer Video-MME value is an input to Eq. (2). The DP allocation (Eq. 4) maximizes answerer accuracy on a separate 1,000-video held-out set; this is validation-based hyperparameter selection, not a fitted quantity being relabeled as a prediction. The reported 6.4-point gain is a measured result against uniform sampling on Video-MME, not a quantity implied by the calibration objective. No equation sets an output equal to an input by construction, no result is justified solely by a self-citation, and there is no imported uniqueness theorem. The only mild caveat is the presence of the Video-MME column in the layer-selection table, which is a potential leakage/transparency concern but not a demonstrated circular reduction.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The central claim rests on an empirical premise about attention concentrativeness, plus several calibration choices: extraction layer, DP allocation, token compression, and important-token sets. None of these are derived from first principles; they are validated empirically on development sets and benchmarks. No new physical or architectural entities are introduced.

free parameters (4)
  • Extraction layer ℓ⋆ per selector backbone = InternVL3.5-2B: L12; Qwen2.5-VL-3B: L23; Qwen3-VL-2B: L14; Qwen3.5-2B: L15
    Chosen by sweeping attention-scoring depth on Charades-STA/NExT-GQA (Table 1, Fig. 4), then frozen; an operating point selected on development data.
  • Duration-bin allocation policy (F(d), P(d)) = DP lookup table under B=32,000
    Selected by maximizing held-out QA accuracy on CG-Bench, NExT-QA, LongVideo-Reason, and LLaVA-Video (Sec. 4.1), subject to budget and monotonicity constraints; effectively a fitted schedule.
  • Selector visual-token budget B = 32,000
    Manually chosen context budget; controls all frame/token allocations.
  • Important-token set T_imp composition = Answer-option tokens for multiple-choice; LLM-extracted entity tokens for open-ended QA
    Task-dependent design choice in Sec. 3.2; no stated criterion beyond these heuristics, and the open-ended variant depends on an unspecified external LLM.
axioms (6)
  • domain assumption Cross-modal attention at a selected layer is a valid query-relevant evidence signal.
    Central empirical premise of DAFS; supported only by validation sweeps and downstream transfer, not derived.
  • domain assumption Uniform subsampling of candidate frames at base 1 FPS preserves query-relevant evidence.
    Used in Sec. 3.3; if false, the selector cannot recover evidence that falls between sampled instants.
  • domain assumption Non-overlapping average pooling of vision tokens to P tokens preserves scoring-relevant spatial evidence.
    Token compression in Sec. 3.3; the paper provides no analysis of information loss from pooling.
  • domain assumption Monotonic allocation constraints (F_i ≤ F_{i+1}, P_i ≥ P_{i+1}) do not exclude the optimal policy.
    Eqs. (6)-(7) in Sec. 3.4 restrict the search space; if the optimal allocation ever violates monotonicity, the calibrated policy is suboptimal.
  • domain assumption An external LLM can extract reliable important tokens for open-ended QA.
    Sec. 3.2 relies on an unspecified LLM for T_imp in open-ended tasks; potential source of variance and cost.
  • standard math The dynamic programming formulation satisfies optimal substructure for the stated discrete optimization.
    Standard DP assumption; the objective values are precomputed from calibration accuracy, so the solver is valid if the objective is additive.

pith-pipeline@v1.3.0-alltime-deepseek · 13104 in / 18033 out tokens · 146579 ms · 2026-08-01T22:34:02.582487+00:00 · methodology

0 comments
read the original abstract

Understanding long videos with multimodal large language models (MLLMs) requires selecting a compact set of frames from thousands of candidates, yet identifying the right frames seemingly requires understanding the video first. We resolve this circular dependency with a simple observation: cross-modal attention at validation-selected extraction layers in MLLMs already provides query-relevant frame evidence without requiring autoregressive generation. We exploit this property to build DAFS (Dynamic Attention-based Budget-aware Frame Selection), a training-free frame selector. A lightweight MLLM selector, even with only 2B parameters, can extract frame-level evidence by converting selected-layer attention into relevance scores through query-conditioned aggregation. This enables cross-frame comparison without autoregressive decoding. To handle the selector's own context constraint, we formulate the joint allocation of candidate pool size and per-frame token budget as a discrete optimization problem solved by dynamic programming. Under a 32-frame budget, our selector improves over uniform sampling by up to 6.4 points on Video-MME and outperforms prior training-based selectors under matched frame budgets, while generalizing across selector and answerer backbones, and across tasks, without retraining.

Figures

Figures reproduced from arXiv: 2607.15689 by Alex Jinpeng Wang, Dongxing Mao, Linjie Li, Ping Yu, Rui Yan, Xiangxi Zheng, Yilin Wang, Yuan Yao, Zhengyuan Yang.

Figure 1
Figure 1. Figure 1: Performance–efficiency overview. Our selector improves long-video understanding while keeping the selec￾tion stage lightweight, illustrating the accuracy–cost trade-off targeted by this work. processing computationally infeasible. Therefore, long video understanding largely reduces to a prob￾lem: selecting a small set of informative frames that retrieves query-relevant visual evidence. Existing approaches … view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of text-guided frame selection paradigms for long video understanding. CLIP-based methods score frames independently with weak query conditioning. MLLM-based autoregressive selectors achieve strong query conditioning but require additional training. Our method extracts cross-modal attention from a small MLLM in a single forward pass for training-free, query-conditioned selection. jointly. In add… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of DAFS. Candidate frames are compressed via dynamic pooling and scored by a small MLLM using validation-selected cross-modal attention. The top-K query-relevant frames are then fed to the backbone model for prediction. Charades-STA and NExT-GQA. Over the top-32 scored frames, Hit@32 is the fraction of queries hitting a ground-truth frame, Cov@32 the covered fraction of the ground-truth segment, a… view at source ↗
Figure 4
Figure 4. Figure 4: Layer-wise attention behavior. Left: temporal-evidence recall on NExT-GQA when using each selector depth as the frame-scoring layer; depth 12 gives the strongest recall. Right: cosine similarity between the frame-level attention-score distributions of each layer and the next layer, showing how the attention pattern changes with depth. sampling and spatial token compression. The pa￾rameters for both are joi… view at source ↗
Figure 5
Figure 5. Figure 5: A Video-MME qualitative case where our selected [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Tool-use generalization. Video-MME accuracy under the preview → retrieve → answer workflow with GPT￾4o. 32-frame budget, the same selector consistently im￾proves each answerer, with gains up to +3.5 points on Video-MME, +4.6 on LVB, +6.8 on MLVU M￾Avg, and +12.4 on MLVU N-QA, so the selected frames transfer beyond a particular answering back￾bone. Generalization to other tasks. On MMBench￾Video, we follow … view at source ↗

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

51 extracted references · 22 linked inside Pith

  1. [1]

    2025 , eprint=

    Qwen2.5-VL Technical Report , author=. 2025 , eprint=

    2025

  2. [2]

    Proceedings of the 2024 conference on empirical methods in natural language processing , pages=

    Video-llava: Learning united visual representation by alignment before projection , author=. Proceedings of the 2024 conference on empirical methods in natural language processing , pages=

    2024

  3. [3]

    Science China Information Sciences , volume=

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

    2025

  4. [4]

    arXiv preprint arXiv:2406.07476 , year=

    Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms , author=. arXiv preprint arXiv:2406.07476 , year=

    Pith/arXiv arXiv

  5. [5]

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

    Chat-univi: Unified visual representation empowers large language models with image and video understanding , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  6. [6]

    arXiv preprint arXiv:2408.03326 , year=

    Llava-onevision: Easy visual task transfer , author=. arXiv preprint arXiv:2408.03326 , year=

    Pith/arXiv arXiv

  7. [7]

    arXiv preprint arXiv:2407.15841 , year=

    Slowfast-llava: A strong training-free baseline for video large language models , author=. arXiv preprint arXiv:2407.15841 , year=

    Pith/arXiv arXiv

  8. [8]

    arXiv preprint arXiv:2408.10188 , year=

    Longvila: Scaling long-context visual language models for long videos , author=. arXiv preprint arXiv:2408.10188 , year=

    Pith/arXiv arXiv

  9. [9]

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

    Video-xl: Extra-long vision language model for hour-scale video understanding , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  10. [10]

    arXiv preprint arXiv:2408.14023 , year=

    Video-ccam: Enhancing video-language understanding with causal cross-attention masks for short and long videos , author=. arXiv preprint arXiv:2408.14023 , year=

    Pith/arXiv arXiv

  11. [11]

    arXiv preprint arXiv:2410.17434 , year=

    Longvu: Spatiotemporal adaptive compression for long video-language understanding , author=. arXiv preprint arXiv:2410.17434 , year=

    Pith/arXiv arXiv

  12. [12]

    arXiv preprint arXiv:2504.02438 , year=

    Scaling video-language models to 10k frames via hierarchical differential distillation , author=. arXiv preprint arXiv:2504.02438 , year=

    Pith/arXiv arXiv

  13. [13]

    arXiv preprint arXiv:2507.01949 , year=

    Kwai keye-vl technical report , author=. arXiv preprint arXiv:2507.01949 , year=

    Pith/arXiv arXiv

  14. [14]

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

    Bolt: Boost large vision-language model without training for long-form video understanding , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  15. [15]

    arXiv preprint arXiv:2502.06428 , year=

    Cos: Chain-of-shot prompting for long video understanding , author=. arXiv preprint arXiv:2502.06428 , year=

    Pith/arXiv arXiv

  16. [16]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Mdp3: A training-free approach for list-wise frame selection in video-llms , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  17. [17]

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

    Adaptive keyframe sampling for long video understanding , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  18. [18]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Q-frame: Query-aware frame selection and multi-resolution adaptation for video-llms , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  19. [19]

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

    Flexible frame selection for efficient video reasoning , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  20. [20]

    arXiv preprint arXiv:2410.03226 , year=

    Frame-voyager: Learning to query frames for video large language models , author=. arXiv preprint arXiv:2410.03226 , year=

    Pith/arXiv arXiv

  21. [21]

    arXiv preprint arXiv:2505.15447 , year=

    Viarl: Adaptive temporal grounding via visual iterated amplification reinforcement learning , author=. arXiv preprint arXiv:2505.15447 , year=

    Pith/arXiv arXiv

  22. [22]

    2025 , eprint=

    ReFoCUS: Reinforcement-guided Frame Optimization for Contextual Understanding , author=. 2025 , eprint=

    2025

  23. [23]

    Advances in Neural Information Processing Systems , volume=

    Self-chained image-language model for video localization and question answering , author=. Advances in Neural Information Processing Systems , volume=

  24. [24]

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

    Generative frame sampler for long video understanding , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  25. [25]

    arXiv preprint arXiv:2506.00318 , year=

    Chain-of-frames: Advancing video understanding in multimodal llms via frame-aware reasoning , author=. arXiv preprint arXiv:2506.00318 , year=

    Pith/arXiv arXiv

  26. [26]

    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=

  27. [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=

  28. [28]

    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=

  29. [29]

    arXiv preprint arXiv:2502.14786 , year=

    Siglip 2: Multilingual vision-language encoders with improved semantic understanding, localization, and dense features , author=. arXiv preprint arXiv:2502.14786 , year=

    Pith/arXiv arXiv

  30. [30]

    2024 , eprint=

    LMMs-Eval: Reality Check on the Evaluation of Large Multimodal Models , author=. 2024 , eprint=

    2024

  31. [31]

    LMMs-Eval: Accelerating the Development of Large Multimodal Models , url =

    Bo Li* and Peiyuan Zhang* and Kaichen Zhang* and Fanyi Pu* and Xinrun Du and Yuhao Dong and Haotian Liu and Yuanhan Zhang and Ge Zhang and Chunyuan Li and Ziwei Liu , publisher =. LMMs-Eval: Accelerating the Development of Large Multimodal Models , url =

  32. [32]

    2025 , eprint=

    K-frames: Scene-Driven Any-k Keyframe Selection for long video understanding , author=. 2025 , eprint=

    2025

  33. [33]

    5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency , author=

    InternVL3. 5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency , author=. arXiv preprint arXiv:2508.18265 , year=

    Pith/arXiv arXiv

  34. [34]

    2024 , howpublished =

    Gemini 1.5 Flash , author =. 2024 , howpublished =

    2024

  35. [35]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Video-chatgpt: Towards detailed video understanding via large vision and language models , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  36. [36]

    arXiv preprint arXiv:2505.09388 , year=

    Qwen3 Technical Report , author=. arXiv preprint arXiv:2505.09388 , year=

    Pith/arXiv arXiv

  37. [37]

    2025 , eprint =

    Qwen3-VL Technical Report , author =. 2025 , eprint =

    2025

  38. [38]

    2024 , eprint=

    Video Instruction Tuning With Synthetic Data , author=. 2024 , eprint=

    2024

  39. [39]

    arXiv preprint arXiv:2410.02713 , year=

    Llava-video: Video instruction tuning with synthetic data , author=. arXiv preprint arXiv:2410.02713 , year=

    Pith/arXiv arXiv

  40. [40]

    arXiv preprint arXiv:2409.12191 , year=

    Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution , author=. arXiv preprint arXiv:2409.12191 , year=

    Pith/arXiv arXiv

  41. [41]

    European conference on computer vision , pages=

    Internvideo2: Scaling foundation models for multimodal video understanding , author=. European conference on computer vision , pages=. 2024 , organization=

    2024

  42. [42]

    arXiv preprint arXiv:2406.14515 , year=

    MMBench-Video: A Long-Form Multi-Shot Benchmark for Holistic Video Understanding , author=. arXiv preprint arXiv:2406.14515 , year=

    Pith/arXiv arXiv

  43. [43]

    arXiv preprint arXiv:2507.06261 , year=

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities , author=. arXiv preprint arXiv:2507.06261 , year=

    Pith/arXiv arXiv

  44. [44]

    arXiv preprint arXiv:2410.21276 , year=

    Gpt-4o system card , author=. arXiv preprint arXiv:2410.21276 , year=

    Pith/arXiv arXiv

  45. [45]

    arXiv preprint arXiv:2403.05530 , year=

    Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context , author=. arXiv preprint arXiv:2403.05530 , year=

    Pith/arXiv arXiv

  46. [46]

    2023 , url=

    GPT-4V(ision) System Card , author=. 2023 , url=

    2023

  47. [47]

    2025 , eprint=

    M-LLM Based Video Frame Selection for Efficient Video Understanding , author=. 2025 , eprint=

    2025

  48. [48]

    2510.04428 , archivePrefix=

    Yuanhao Zou and Shengji Jin and Andong Deng and Youpeng Zhao and Jun Wang and Chen Chen , year=. 2510.04428 , archivePrefix=

    arXiv

  49. [49]

    2603.25072 , archivePrefix=

    Junpeng Ma and Sashuai Zhou and Guanghao Li and Xin Gao and Yue Cao and Hengyu Zeng and Yuxiang Yan and Zhibin Wang and Jun Song and Bo Zheng and Shanghang Zhang and Jian Pu , year=. 2603.25072 , archivePrefix=

    arXiv

  50. [50]

    2602.22932 , archivePrefix=

    Wenhui Tan and Xiaoyi Yu and Jiaze Li and Yijing Chen and Jianzhong Ju and Zhenbo Luo and Ruihua Song and Jian Luan , year=. 2602.22932 , archivePrefix=

    arXiv

  51. [51]

    arXiv preprint arXiv:2303.08774 , year=

    Gpt-4 technical report , author=. arXiv preprint arXiv:2303.08774 , year=

    Pith/arXiv arXiv