Pith. sign in

REVIEW 3 major objections 5 minor 42 references

Routing each video question as global or local, then sampling frames differently, lifts long-video LVLM accuracy without training.

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 · grok-4.5

2026-07-31 06:35 UTC pith:VPRXSAHJ

load-bearing objection Solid training-free sampler with real matched gains over uniform; the dual-branch “router” story is thinner than the abstract sells, and baseline fairness on scorers is the main open question. the 3 major comments →

arxiv 2607.28463 v1 pith:VPRXSAHJ submitted 2026-07-30 cs.CV

VisualRouter: Query-Grounded Visual Sampling for Long Video Understanding

Haiyue Zhang , Yi Bin , Xun Jiang , Zeyu Ma , Duo Peng , Guoqing Wang , Yang Yang , Heng Tao Shen This is my paper
classification cs.CV
keywords long video understandingvisual samplingquery-grounded routinglarge vision-language modelsevent-aware frame selectiontraining-freedeterminantal point process
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.

Long videos overwhelm vision-language models because too many frames become too many tokens. Most frame pickers either chase query similarity and pile up redundant shots, or use one fixed rule for every question. This paper argues that the right fix is to first decide what kind of evidence the question needs. Global questions (plot, genre, overall summary) need broad temporal coverage mixed with some relevant frames. Local questions (a specific object, action, or event) need the budget spent on the right event segments, with diversity inside those segments so near-duplicate frames do not waste slots. VisualRouter does that routing with the model itself, then runs a hybrid sampler for global queries and an event-partition, allocate, and diversity-select pipeline for local ones. It is training-free and drops in front of existing models. Under matched backbones and frame counts it beats uniform sampling and several strong training-free baselines on Video-MME, LongVideoBench, and MLVU, with the largest lifts when the frame budget is tight.

Core claim

Query-grounded routing plus branch-specific sampling is enough to improve long-video understanding: classify each query as global or local, then apply a relevance-coverage hybrid for global queries and an event-aware pipeline (partition, allocate by segment relevance and length, select inside segments with a query-aware k-DPP) for local queries. That design consistently raises accuracy over uniform sampling and over fixed relevance-, coverage-, or diversity-only samplers under the same LVLM and same number of frames.

What carries the argument

VisualRouter: a training-free query gate that labels each question global or local, then either merges uniform coverage with Top-K relevance (global) or runs event partitioning from fused semantic/visual boundary scores, weighted segment frame allocation, and intra-event k-DPP selection balancing relevance and diversity (local).

Load-bearing premise

The method assumes that BLIP2 relevance scores, DINOv2/BLIP2 boundary cues, and a prompt that labels queries global versus local (with all LongVideoBench items forced local) are good enough stand-ins for where the true answer evidence lives.

What would settle it

Hold the LVLM, prompts, and frame count fixed; replace VisualRouter’s routing and local pipeline with uniform or Top-K only on the same Video-MME, LongVideoBench, and MLVU splits. If the reported gains (for example +5.2%, +7.7%, +11.6% with Qwen2.5-VL-7B at 16 frames) disappear or reverse, the central claim fails.

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

If this is right

  • Long-video LVLMs can gain several points of accuracy from a plug-in preprocessor without retraining or changing architecture.
  • A single fixed sampling policy is suboptimal; global narrative questions and local grounded questions need different frame budgets.
  • Gains are largest under tight frame budgets (e.g., 8 frames), so routing matters most when context is scarce.
  • The same routing-plus-event pipeline transfers across multiple LVLM families and scales from 3B to 72B under matched settings.

Where Pith is reading between the lines

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

  • If routing quality is the bottleneck, a cheap dedicated global/local classifier might match or beat using the full LVLM as the gate while cutting preprocess cost.
  • Event-aware allocation could transfer to other long-context modalities (audio chapters, document sections) where evidence is sparse and queries split into overview versus pinpoint.
  • Failures will cluster where relevance scorers miss OCR, fine counting, or abstract narrative cues—suggesting scorer choice as the next ablation lever beyond the sampler itself.

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 / 5 minor

Summary. The paper proposes VisualRouter, a training-free plug-and-play preprocessor for long-video LVLMs. It routes each query as global or local via an LVLM prompt, then applies either a relevance–coverage hybrid sampler (global) or a three-stage event-aware pipeline—boundary-based event partitioning, segment-level frame allocation via a relevance–length weight, and intra-event k-DPP selection (local). Candidate frames are scored with BLIP2-ITM and DINOv2 features at 1 FPS. Under matched LVLM backbones and frame budgets, the method reports consistent gains over uniform sampling (e.g., +5.2% Video-MME, +7.7% LongVideoBench, +11.6% MLVU with Qwen2.5-VL-7B at 16 frames) and competitive or better results versus reproduced Top-K, BOLT, AKS, and WFS-SB across several backbones (Tables I–II), with supporting ablations (Tables IV–V) and multi-scale/frame-budget checks (Fig. 4, Table III).

Significance. If the matched-setting gains hold under fully controlled baselines, the work is a useful systems contribution: a training-free, model-agnostic sampler that improves several strong open LVLMs without architectural change. The dual-branch design and the event-aware local pipeline (partition → allocate → k-DPP) are clearly specified and ablated, and the multi-backbone, multi-benchmark evaluation is stronger than many sampling papers. Code availability is stated. The practical value is real even if part of the lift comes from a strong relevance scorer rather than routing alone; the field needs reliable frame selection under tight budgets. Credit is due for systematic ablations of gating, allocation, DPP, scorers, and boundary fusion, and for reporting gains across model scales up to 72B.

major comments (3)
  1. [Table II, §IV-B.2, Table V] Table II and §IV-B.2 claim outperformance of training-free samplers “under the same setting,” with BOLT/AKS/WFS-SB marked as reproduced (†). The manuscript never states that those reproductions used the same 1 FPS candidate pool and the same BLIP2-ITM relevance scores as VisualRouter/Top-K. Table V shows that scorer choice alone moves LVB from 56.5 (CLIP) to 63.7 (BLIP2-ITM) inside the authors’ pipeline—on the order of the headline +7.7. Without an explicit matched-scorer protocol (or a controlled re-run), gains over published methods are confounded with scorer quality. Please document the exact candidate rate, features, and relevance model used for each † baseline, or re-evaluate all methods with a shared scorer and candidate set.
  2. [§III-C, Table IV, Abstract, §I contributions] §III-C hard-routes every LongVideoBench sample to the local branch, so the reported +7.7 LVB gain cannot be attributed to global/local query gating. Table IV further shows local-only at 61.8/63.7/67.4 vs full VisualRouter at 62.3/63.7/67.9—within ~0.5 points on the mixed benchmarks. The abstract and contribution bullets foreground “query-grounded” routing as the central design; the evidence instead points to the event-aware local sampler (plus BLIP2-ITM) as the main driver. Please (i) report gating accuracy or agreement on Video-MME/MLVU, (ii) give the global/local split counts, and (iii) revise claims so that credit matches the ablations—routing as a small optional switch, local event-aware selection as the primary mechanism.
  3. [§III-D, Eq. (10), Fig. 4, Table IV] §III-D.2 and Eq. (10): the segment weight w_m = sqrt(smax_m · smean_m) · sqrt(ℓ_m) and the global 50/50 coverage–relevance split, EMA rate ρ, fusion (λ_s, λ_v)=(0.6,0.4), and DPP exponent γ are free design choices with limited sensitivity analysis. Table IV replaces allocation with equal split and sees a clear LVB drop (63.7→60.1), which is good, but there is no sweep on λ, γ, ρ, or the global split ratio. A short sensitivity table (or fixed defaults justified once) is needed so readers can judge robustness of the “same setting” gains, especially under the limited frame budgets emphasized in Fig. 4.
minor comments (5)
  1. [Fig. 1] Fig. 1 category gains (OCR +12.95%, counting +7.09%) are strong but only for one backbone/frame setting; a sentence on whether the same category pattern holds for other LVLMs would help.
  2. [Algorithm 1, §III-D.2] Algorithm 1 references GreedyDPP and quality map q_i=(s_itm_i)^γ but γ is never numerically specified in the main text; add the default value used in all tables.
  3. [§III-A, Eq. (1)] Eq. (1) writes K·m+ℓ_Q ≤ L; m is never measured for the evaluated LVLMs. Either drop the unused constraint or report typical token counts.
  4. [§II-B] Related work cites concurrent/arXiv sampling methods densely; a short explicit “difference from AKS/BOLT/WFS-SB” paragraph in §II-B would clarify novelty beyond routing.
  5. [§III-A, Fig. 1] Typos/style: “F ormulation” spacing in §III-A; “Info. Synopsis” in Fig. 1; arXiv IDs in references are fine but ensure camera-ready consistency for venue names.

Circularity Check

0 steps flagged

No circularity: empirical plug-and-play sampler evaluated on external benchmarks; claims are measured accuracy, not derived identities.

full rationale

VisualRouter is a training-free engineering pipeline (query gate → global hybrid or local event-aware selection using BLIP2-ITM/DINOv2 signals and k-DPP). The load-bearing claims are comparative accuracies on Video-MME, LongVideoBench, and MLVU under matched LVLM and frame budgets (Abstract; Tables I–II). Those numbers are external benchmark measurements, not quantities obtained by fitting a parameter and re-reporting it, nor by defining the objective in terms of the reported metric. Method choices (prompt-defined GLOBAL/LOCAL labels, λ_s/λ_v, hard-routing LVB to local, γ in q_i=s^γ) are hyperparameters of the sampler; they do not make the accuracy gains true by construction. Related-work citations are standard external methods and models, not self-authored uniqueness theorems that force the result. Skeptical concerns about scorer confounding or weak attribution to gating are correctness/fairness issues, not circular derivation. Score 0 with empty steps is appropriate.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 2 invented entities

Load-bearing content is mostly engineering assumptions and hand-chosen knobs, not deep physical axioms. The claim rests on off-the-shelf scorers, a prompt taxonomy of queries, fixed fusion/allocation rules, and standard benchmark protocols. No new particles or forces; the ‘entities’ are methodological constructs.

free parameters (5)
  • boundary fusion weights λs, λv = λs=0.6, λv=0.4
    Set to 0.6 and 0.4 for semantic vs visual boundary scores without a principled derivation; directly affect event partitions.
  • global branch coverage/relevance split = ≈50/50 of K
    Allocates floor(K/2) frames to uniform coverage and the rest to Top relevance; a design choice that shapes global-query behavior.
  • DPP quality exponent γ = γ>1 (exact default not stated)
    q_i = (s_itm_i)^γ with γ>1 sharpens relevance inside k-DPP; value not fully specified as a single reported default in the main text.
  • EMA update rate ρ = unspecified numeric default
    Controls semantic-drift reference for event boundaries; chosen for stability, not derived.
  • candidate sampling rate = 1 FPS
    All videos reduced to 1 FPS candidates before scoring/selection; changes the discrete search space.
axioms (6)
  • domain assumption Queries can be usefully partitioned into global vs local by presence of visual anchors in question or options.
    Sec. III-C and Fig. 3 define the routing taxonomy that chooses the entire sampling branch.
  • domain assumption BLIP2-ITM scores are adequate proxies for query-frame semantic relevance for both Top-K and segment weights.
    Sec. III-B Eq. (3); Table V compares scorers but the pipeline assumes ITM is good enough for allocation and DPP quality.
  • domain assumption Fused semantic drift and DINOv2 visual change peaks mark event boundaries that matter for evidence allocation.
    Sec. III-D event partitioning Eqs. (5)–(7); ablation in Table V.
  • ad hoc to paper All LongVideoBench items should be routed to the local branch without per-query gating.
    Explicit exception in Sec. III-C; bypasses the general router on one of three main benchmarks.
  • domain assumption Fixed-size k-DPP with greedy log-det maximization sufficiently trades relevance and diversity inside segments.
    Sec. III-D intra-event selection Eqs. (12)–(14); standard ML approximation assumed adequate.
  • domain assumption Matched frame count K and frozen LVLM weights make accuracy deltas attributable to the sampler.
    Evaluation protocol in Sec. IV-A; core fairness claim for Tables I–II.
invented entities (2)
  • VisualRouter dual-branch query gate independent evidence
    purpose: Route each query to global hybrid sampling or local event-aware sampling.
    Operational framework introduced in Sec. III; not a physical entity, but the paper’s central postulated mechanism.
  • Segment importance weight w_m = sqrt(smax_m * smean_m) * sqrt(ℓ_m) no independent evidence
    purpose: Allocate frame budget across event segments from peak/mean relevance and length.
    Ad hoc combination in Eq. (10); justified by intuition, validated only via ablation.

pith-pipeline@v1.2.0-daily-grok45 · 21769 in / 3882 out tokens · 78902 ms · 2026-07-31T06:35:00.977702+00:00 · methodology

0 comments
read the original abstract

Large vision-language models (LVLMs) have achieved significant progress in video understanding, yet understanding long videos remains challenging due to the large number of visual tokens and limited context windows. Visual sampling provides a practical solution by selecting an informative subset of frames. However, existing methods typically either rely on relevance-aware sampling, leading to redundant frame selection and insufficient temporal coverage, or adopt a fixed sampling strategy regardless of query type. In this paper, we propose VisualRouter, a training-free and plug-and-play framework for query-grounded visual sampling. VisualRouter first classifies each query as either global or local and then applies the corresponding sampling strategy. For global queries, it employs a relevance-coverage hybrid strategy that preserves temporal coverage while retaining query-relevant visual evidence. For local queries, it adopts an event-aware frame selection strategy that performs event partitioning, segment-level frame allocation, and intra-event frame selection, jointly balancing relevance, coverage, and diversity with a limited number of input frames. Experiments show that VisualRouter consistently improves multiple LVLMs over uniform sampling, achieving gains of 5.2%, 7.7%, and 11.6% on Video-MME, LongVideoBench, and MLVU with Qwen2.5-VL-7B, and outperforming existing training-free visual sampling methods under the same setting.

Figures

Figures reproduced from arXiv: 2607.28463 by Duo Peng, Guoqing Wang, Haiyue Zhang, Heng Tao Shen, Xun Jiang, Yang Yang, Yi Bin, Zeyu Ma.

Figure 1
Figure 1. Figure 1: Accuracy(%) of visual sampling methods across six question cate [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of VisualRouter. Each query is first routed to either the global or local branch. The global branch combines relevance-aware selection with [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Prompt design for query gating. The prompt determines whether a [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Effect of the number of input frames. Dark bars show the accuracy of uniform sampling, while light extensions show the accuracy gains achieved by [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative examples from Video-MME comparing uniform sampling with VisualRouter. In the examples shown, VisualRouter selects frames from [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] 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

42 extracted references · 7 linked inside Pith

  1. [1]

    Video-ChatGPT: Towards detailed video understanding via large vision and language models,

    M. Maaz, H. Rasheed, S. Khan, and F. Khan, “Video-ChatGPT: Towards detailed video understanding via large vision and language models,” inProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, 2024, pp. 12 585–12 602

    2024

  2. [2]

    LLaV A-video: Video instruction tuning with synthetic data,

    Y . Zhang, J. Wu, W. Li, B. Li, Z. Ma, Z. Liu, and C. Li, “LLaV A-video: Video instruction tuning with synthetic data,”Transactions on Machine Learning Research, 2025

    2025

  3. [3]

    LLaV A-onevision: Easy visual task transfer,

    B. Li, Y . Zhang, D. Guoet al., “LLaV A-onevision: Easy visual task transfer,”Transactions on Machine Learning Research, 2025

    2025

  4. [4]

    Qwen2.5-VL technical report,

    S. Bai, K. Chen, X. Liuet al., “Qwen2.5-VL technical report,”arXiv preprint arXiv:2502.13923, 2025

    Pith/arXiv arXiv 2025

  5. [5]

    Video- LLaV A: Learning united visual representation by alignment before projection,

    B. Lin, Y . Ye, B. Zhu, J. Cui, M. Ning, P. Jin, and L. Yuan, “Video- LLaV A: Learning united visual representation by alignment before projection,”arXiv preprint arXiv:2311.10122, 2023

    Pith/arXiv arXiv 2023

  6. [6]

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

    C. Fu, Y . Dai, Y . Luoet al., “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, 2025, pp. 24 108–24 118

    2025

  7. [7]

    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,” inAdvances in Neural Information Processing Systems, vol. 37, 2024, pp. 28 828– 28 857

    2024

  8. [8]

    MLVU: Benchmarking multi-task long video understanding,

    J. Zhou, Y . Shu, B. Zhaoet al., “MLVU: Benchmarking multi-task long video understanding,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 13 691–13 701

    2025

  9. [9]

    SlowFast-LLaV A-1.5: A family of token- efficient video large language models for long-form video understand- ing,

    M. Xu, M. Gao, S. Liet al., “SlowFast-LLaV A-1.5: A family of token- efficient video large language models for long-form video understand- ing,” inConference on Language Modeling, 2025

    2025

  10. [10]

    Long context transfer from language to vision,

    P. Zhang, K. Zhang, B. Liet al., “Long context transfer from language to vision,”Transactions on Machine Learning Research, 2025

    2025

  11. [11]

    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,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 29 118–29 128

    2025

  12. [12]

    Q-Frame: Query-aware frame selection and multi-resolution adaptation for video-LLMs,

    S. Zhang, J. Yang, J. Yin, Z. Luo, and J. Luan, “Q-Frame: Query-aware frame selection and multi-resolution adaptation for video-LLMs,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 22 056–22 065

    2025

  13. [13]

    BOLT: Boost large vision- language model without training for long-form video understanding,

    S. Liu, C. Zhao, T. Xu, and B. Ghanem, “BOLT: Boost large vision- language model without training for long-form video understanding,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 3318–3327

    2025

  14. [14]

    MDP3: A training-free approach for list- wise frame selection in video-LLMs,

    H. Sun, S. Lu, H. Wanget al., “MDP3: A training-free approach for list- wise frame selection in video-LLMs,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 24 090–24 101

    2025

  15. [15]

    KeyVideoLLM: Towards large-scale video keyframe selection,

    H. Liang, J. Li, T. Baiet al., “KeyVideoLLM: Towards large-scale video keyframe selection,”arXiv preprint arXiv:2407.03104, 2024

    Pith/arXiv arXiv 2024

  16. [16]

    MaxInfo: A training-free key-frame selection method using maximum volume for enhanced video understanding,

    P. Li, I. Abdullaeva, A. Gambashidze, A. Kuznetsov, and I. Oseledets, “MaxInfo: A training-free key-frame selection method using maximum volume for enhanced video understanding,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2026, pp. 7198–7207

    2026

  17. [17]

    AdaRD-Key: Adaptive relevance- diversity keyframe sampling for long-form video understanding,

    X. Zhang, Z. Wu, Z. Liet al., “AdaRD-Key: Adaptive relevance- diversity keyframe sampling for long-form video understanding,”arXiv preprint arXiv:2510.02778, 2025

    arXiv 2025

  18. [18]

    InternVideo2.5: Empowering video MLLMs with long and rich context modeling,

    Y . Wang, X. Li, Z. Yanet al., “InternVideo2.5: Empowering video MLLMs with long and rich context modeling,”arXiv preprint arXiv:2501.12386, 2025

    Pith/arXiv arXiv 2025

  19. [19]

    VideoLLaMA 3: Frontier multimodal foundation models for image and video understanding,

    B. Zhang, K. Li, Z. Chenget al., “VideoLLaMA 3: Frontier multimodal foundation models for image and video understanding,”arXiv preprint arXiv:2501.13106, 2025

    Pith/arXiv arXiv 2025

  20. [20]

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

    X. Shen, Y . Xiong, C. Zhaoet al., “LongVU: Spatiotemporal adaptive compression for long video-language understanding,” inProceedings of the 42nd International Conference on Machine Learning, vol. 267, 2025, pp. 54 582–54 599

    2025

  21. [21]

    Flexible frame selection for efficient video reasoning,

    S. Buch, A. Nagrani, A. Arnab, and C. Schmid, “Flexible frame selection for efficient video reasoning,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 29 071–29 082

    2025

  22. [22]

    Frame-V oyager: Learning to query frames for video large language models,

    S. Yu, C. Jin, H. Wanget al., “Frame-V oyager: Learning to query frames for video large language models,” inInternational Conference on Learning Representations, 2025

    2025

  23. [23]

    M-LLM based video frame selection for efficient video understanding,

    K. Hu, F. Gao, X. Nieet al., “M-LLM based video frame selection for efficient video understanding,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 13 702–13 712

    2025

  24. [24]

    Self-chained image-language model for video localization and question answering,

    S. Yu, J. Cho, P. Yadav, and M. Bansal, “Self-chained image-language model for video localization and question answering,” inAdvances in Neural Information Processing Systems, vol. 36, 2023

    2023

  25. [25]

    Generative frame sampler for long video understanding,

    L. Yao, H. Wu, K. Ouyanget al., “Generative frame sampler for long video understanding,” inFindings of the Association for Computational Linguistics: ACL 2025, 2025, pp. 17 900–17 917

    2025

  26. [26]

    K-frames: Scene-driven any- k keyframe selection for long video understanding,

    Y . Yao, Y . Yun, J. Wanget al., “K-frames: Scene-driven any- k keyframe selection for long video understanding,”arXiv preprint arXiv:2510.13891, 2025

    arXiv 2025

  27. [27]

    Event-anchored frame selection for effective long-video understanding,

    W. Chen, Y . Luo, Y . Zenget al., “Event-anchored frame selection for effective long-video understanding,”arXiv preprint arXiv:2603.00983, 2026

    arXiv 2026

  28. [28]

    Wavelet-based frame selection by detecting semantic boundary for long video understanding,

    W. Chen, Y . Zeng, Y . Luoet al., “Wavelet-based frame selection by detecting semantic boundary for long video understanding,” inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 24 052–24 061

    2026

  29. [29]

    The use of MMR, diversity-based rerank- ing for reordering documents and producing summaries,

    J. Carbonell and J. Goldstein, “The use of MMR, diversity-based rerank- ing for reordering documents and producing summaries,” inProceedings of the 21st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, 1998, pp. 335–336

    1998

  30. [30]

    BLIP-2: Bootstrapping language- image pre-training with frozen image encoders and large language mod- els,

    J. Li, D. Li, S. Savarese, and S. Hoi, “BLIP-2: Bootstrapping language- image pre-training with frozen image encoders and large language mod- els,” inProceedings of the 40th International Conference on Machine Learning, vol. 202, 2023, pp. 19 730–19 742

    2023

  31. [31]

    DINOv2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanniet al., “DINOv2: Learning robust visual features without supervision,”Transactions on Machine Learning Research, 2024

    2024

  32. [32]

    Determinantal point processes for machine learning,

    A. Kulesza and B. Taskar, “Determinantal point processes for machine learning,”F oundations and Trends in Machine Learning, vol. 5, no. 2–3, pp. 123–286, 2012

    2012

  33. [33]

    k-DPPs: Fixed-size determinantal point processes,

    A. Kulesza and B. Taskar, “k-DPPs: Fixed-size determinantal point processes,” inProceedings of the 28th International Conference on Machine Learning, 2011, pp. 1193–1200

    2011

  34. [34]

    Fast greedy MAP inference for determinantal point process to improve recommendation diversity,

    L. Chen, G. Zhang, and E. Zhou, “Fast greedy MAP inference for determinantal point process to improve recommendation diversity,” in Advances in Neural Information Processing Systems, vol. 31, 2018, pp. 5627–5638

    2018

  35. [35]

    LMMs-Eval: Reality check on the evaluation of large multimodal models,

    K. Zhang, B. Li, P. Zhanget al., “LMMs-Eval: Reality check on the evaluation of large multimodal models,” inFindings of the Association for Computational Linguistics: NAACL 2025, 2025, pp. 881–916

    2025

  36. [36]

    Qwen3-VL technical report,

    S. Bai, Y . Cai, R. Chenet al., “Qwen3-VL technical report,”arXiv preprint arXiv:2511.21631, 2025

    Pith/arXiv arXiv 2025

  37. [37]

    InternVL3: Exploring advanced training and test-time recipes for open-source multimodal models,

    J. Zhu, W. Wang, Z. Chenet al., “InternVL3: Exploring advanced training and test-time recipes for open-source multimodal models,”arXiv preprint arXiv:2504.10479, 2025

    Pith/arXiv arXiv 2025

  38. [38]

    LongVILA: Scaling long-context visual language models for long videos,

    Y . Chen, F. Xue, D. Liet al., “LongVILA: Scaling long-context visual language models for long videos,” inInternational Conference on Learning Representations, 2025

    2025

  39. [39]

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

    Y . Shu, Z. Liu, P. Zhanget al., “Video-XL: Extra-long vision language model for hour-scale video understanding,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025, pp. 26 160–26 169

    2025

  40. [40]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacyet al., “Learning transferable visual models from natural language supervision,” inProceedings of the 38th International Conference on Machine Learning, vol. 139, 2021, pp. 8748–8763

    2021

  41. [41]

    Sigmoid loss for language image pre-training,

    X. Zhai, B. Mustafa, A. Kolesnikov, and L. Beyer, “Sigmoid loss for language image pre-training,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 11 975–11 986

    2023

  42. [42]

    BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation,

    J. Li, D. Li, C. Xiong, and S. Hoi, “BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation,” inProceedings of the 39th International Conference on Machine Learn- ing, vol. 162, 2022, pp. 12 888–12 900

    2022