Pith. sign in

REVIEW 5 major objections 5 minor 1 cited by

Threading Keyframe with Narratives: MLLMs as Strong Long Video Comprehenders

T0 review · 5 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Nar-KFC threads query-relevant, diverse keyframes with short captions of skipped frames, and this training-free input rearrangement lifts existing MLLMs on long-video benchmarks, including LLaVA-Video 7B from 55.9% to 61.6% on Video-MME…

desk verdict Solid training-free method that likely helps long-video QA, but the headline gains are inflated by tuning on the eval benchmark and the 'consistent' claim is too strong. read the letter →

arxiv 2505.24158 v1 pith:K4JYYSFU submitted 2025-05-30 cs.CV

Bo Fang , Wenhao Wu , Qiangqiang Wu , Yuxin Song , Antoni B. Chan This is my paper
classification cs.CV
keywords longvideounderstandingmultimodallargelanguagemodelskeyframeselectionintegerquadraticprogramminggreedysearchnarrativethreadingtraining-freequestionanswering
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

Nar-KFC is a training-free module that lets off-the-shelf multimodal LLMs (MLLMs) answer questions about hour-long videos by replacing uniform frame sampling with an optimized set of keyframes threaded by short captions of the skipped frames. The paper argues that query-relevance alone selects redundant nearby frames, while uniform sampling misses key moments, so it casts keyframe selection as an integer quadratic program — a discrete optimization that picks K frames to maximize a combined relevance-diversity score — then solves it with a custom greedy search. Short captions of non-keyframes are inserted between the selected keyframes in their true temporal order, producing a compact interleaved image-text representation that restores temporal continuity. Across Video-MME, LongVideoBench, and MLVU, this improves four popular MLLMs; for example, LLaVA-Video 7B rises from 55.9% to 61.6% on Video-MME without subtitles. The work matters because it suggests that current MLLMs can become strong long-video comprehenders without any additional training, simply by choosing and arranging inputs more intelligently.

What carries the argument

The central mechanism is the interleaved keyframe-narrative sequence. KFC selects K frames by optimizing the integer quadratic objective $\mathbf{x}^T S \mathbf{x}$, where each entry $S(i,j) = \mathrm{sim}(f_i,q) + \alpha \exp(-\mathrm{sim}(f_i,f_j))$ combines query relevance with frame diversity, and a greedy search with SVD (singular value decomposition) low-rank preprocessing, downsampling, and k-nearest-neighbor refinement finds a near-optimal solution in $O(NK)$ time. Nar-KFC then generates short frame-level captions with an off-the-shelf captioner and inserts them between keyframes according to true temporal order, producing the input $M(\{f_{y_1}, c_{y_1+\Delta}, \ldots, f_{y_K}\}, q)$, where $\Delta$ controls narrative density. This interleaved structure is what carries the argument: it jointly optimizes query relevance, content diversity, and temporal continuity within the MLLM's context limit.

What would settle it

A decisive experiment would use a long-video QA set where each question requires a fact that appears only in a non-keyframe and is deliberately left out of the narrative captions. If Nar-KFC still answers correctly, the claim that threading restores temporal continuity is not doing the work; if it fails, the gain must come from the captions themselves rather than from the interleaved structure. The paper's own ordering ablation, where placing all keyframes before or after the narratives costs about 1% accuracy, is a weaker version of this test.

Watch

Extended reading notes

Core claim

The paper's central claim is that long-video understanding with current MLLMs is limited less by model capacity than by how frames are sampled and presented. Nar-KFC threads KFC-selected keyframes with interleaved textual narratives of non-keyframes, constructing a temporally continuous, compact representation of the video. The authors show consistent gains across four MLLMs and three benchmarks, with the strongest result being LLaVA-Video 7B improving from 55.9% to 61.6% on Video-MME without subtitles. The claim is that this works in a training-free, plug-and-play manner, treating existing MLLMs as already capable of long-video comprehension once the input is structured as a coherent story of frames and captions.

Load-bearing premise

The method's gains rest on the premise that MLLMs, even without training on interleaved image-text-image inputs, can read a sequence of keyframes and interleaved captions as one continuous temporal story rather than as unrelated fragments.

Editorial extensions

If this is right

  • With only 8 keyframes, Nar-KFC improves all four tested MLLMs on Video-MME, LongVideoBench, and MLVU; LLaVA-Video reaches 61.6% on Video-MME without subtitles, beating larger models and models that use 256 frames.
  • Narratives add the most when the number of keyframes is small (4 or 8) and on tasks that need temporal continuity, such as MLVU's plotQA and topic categories, while KFC alone already captures query-relevant needle and count frames.
  • A lightweight 2B captioner delivers nearly the same performance as a 72B captioner, with less than 1% gap, so the approach stays cheap in tokens and compute.
  • The greedy search closely matches the exact integer-programming solution while running in $O(NK)$ time, making the optimization practical for hour-long videos.
  • Because the module is training-free and model-agnostic, the same keyframe selection and narrative threading can be dropped onto future MLLMs without retraining.

Reading between the lines

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

  • Editorial inference: the same interleaved-representation idea could be carried over to temporal grounding and dense video captioning, where correct ordering matters more than in multiple-choice QA; the paper's MLVU topic and plotQA gains are the first hint.
  • Editorial inference: if the paper's premise about interleaved inputs is correct, an MLLM explicitly trained on interleaved keyframe-caption sequences should improve further; Appendix A's stated limitation points directly to that experiment.
  • Editorial inference: the small accuracy dips with KFC alone on EgoSchema and NExTQA suggest that the optimal keyframe strategy depends on video length, and a duration-aware switch between uniform sampling and adaptive selection is a natural extension the paper does not test.
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

5 major / 5 minor

Summary. The paper proposes Nar-KFC, a training-free module for long-video understanding with MLLMs. It first selects K keyframes by optimizing a score that combines CLIP-based query relevance and frame diversity, formulating the selection as an integer quadratic program and solving it with a custom greedy search (KFC). It then inserts short captions of non-keyframes, generated by a lightweight captioner, between the selected keyframes in temporal order, yielding an interleaved visual-textual representation (Nar-KFC). Experiments on Video-MME, LongVideoBench, and MLVU with InternVL2, Qwen2-VL, LLaVA-OneVision, and LLaVA-Video report accuracy improvements over uniform sampling, with a headline gain from 55.9% to 61.6% on Video-MME (no subtitles) for LLaVA-Video-7B. The paper also provides ablations on components, number of narratives, captioner choice, and greedy-search sub-steps, plus additional results on EgoSchema and NExTQA in the appendix.

Significance. If the reported gains are robust, Nar-KFC is a practical and conceptually clean contribution: it is training-free, compatible with several MLLMs, and its O(NK) greedy search is much cheaper than the exponential IQP formulation. The idea of threading sparse keyframes with temporally ordered captions is intuitive and well motivated, and the paper includes extensive ablations and an honest discussion of limitations in Appendix A. However, the central claim of consistent, significant improvement is currently under-supported: several hyperparameters are selected on the same benchmarks used for the headline results, the paper's own Table 1 contains drops for Qwen2-VL, and no statistical significance or variance is reported. These issues make the current evidence fragile and the headline numbers potentially inflated, so the significance of the contribution is real but not yet demonstrated to the standard claimed.

major comments (5)
  1. [Sec. 4.2, Fig. 4 and Appendix D.3, Table D.5] The default number of narratives (210) and the refinement window k=2 are selected after observing Video-MME accuracy on the same benchmark used for the headline result (LLaVA-Video 55.9 to 61.6 on Video-MME no subtitles). Other knobs, such as alpha=1, K=8, SVD rank N/4, and downsampling to 128x128, are fixed without a held-out split. Since several interacting hyperparameters are effectively tuned on the test set, the reported gains may be inflated. Please add a validation split or pre-specify the full configuration before evaluation, and report how the main results vary across reasonable choices of these hyperparameters.
  2. [Sec. 4.1, Table 1] The claim of consistent improvements is contradicted by the paper's own numbers: for Qwen2-VL, Nar-KFC drops relative to KFC on LongVideoBench (54.6 to 53.6 overall, and 47.9 to 46.3 on the 15-60 minute subset) and on Video-MME with subtitles (58.3 to 57.9). The statement that KFC and Nar-KFC deliver "consistent and significant gain over four baselines across three long-video benchmarks" therefore needs to be revised to a per-model, per-benchmark discussion, or supported by an explanation of why narratives hurt in these cases.
  3. [Sec. 4.1, 'Comparisons with varying number of keyframes' and Fig. 3] The figure omits Qwen2-VL results because of memory overflow, yet the text states that "both KFC and Nar-KFC consistently outperform uniform sampling" across all settings. This claim is not supported for Qwen2-VL. Please include Qwen2-VL results where feasible, or explicitly restrict the claim to the models shown and add a discussion of the memory limitation.
  4. [Sec. 4, 'Evaluation Models'] The four MLLM baselines are re-implemented with VLMEvalKit, while numbers for prior methods are taken from other toolkits, and no error bars or repeated-run statistics are reported. Since many improvements are around one percentage point, it is possible that some differences are not statistically significant. Please report variance or at least multiple runs, and ensure that all numbers in Table 1 and the appendix are obtained under a common evaluation protocol.
  5. [Appendix A] The paper honestly acknowledges that MLLMs "are not trained with such input formats" for interleaved keyframes and narratives, which is a genuine limitation of the central mechanism. The positive empirical results partially mitigate this concern, but the paper should either provide a targeted experiment comparing interleaved versus concatenated inputs, or soften the claim that Nar-KFC reconstructs a coherent and continuous representation that MLLMs can naturally process.
minor comments (5)
  1. [Appendix C.2 and Table D.5] There are typos: "perfom" in Appendix C.2 should be "perform", "Meidum" in Table D.5 should be "Medium", and "trangular" in Appendix D.1 should be "triangular".
  2. [Fig. 4] The x-axis label "Number of narrations" is inconsistent with the text, which uses "narratives"; please unify the terminology throughout the paper.
  3. [Sec. 4.1, MLVU description] The benchmark description mentions "M-avg" but the acronym is not defined in the main text; please define it when first used.
  4. [Eq. (4)] The notation c_{y1+Delta} is not precisely defined; please clarify that it denotes captions for frames at intermediate timestamps between consecutive keyframes, and specify how the uniform interval Delta is chosen.
  5. [Sec. 4.1, first paragraph] The sentence "We conduct comprehensive comparisons between our approach and several recent MLLMs and VideoLLMs with default K=8 frames" should explicitly state that this K is the number of keyframes, not the total number of input items when narratives are included.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Nar-KFC is an empirical selection-and-narration pipeline whose benchmark gains are measured, not derived from fitted inputs or self-citations.

full rationale

The paper's derivation chain is an optimization formulation (Eqs. 1-3) plus a greedy approximation (Alg. 1) and an interleaving construction (Eq. 4). Nothing in the objective is defined in terms of the benchmark answers; query-relevance and frame-diversity scores are computed from CLIP embeddings, and the selected keyframes with narratives are consumed by a frozen MLLM whose output is an empirical measurement. No fitted parameter is renamed as a prediction, and no load-bearing claim rests on a self-citation: the only co-author citation (Ref. [72], Dense Connector) appears in the introduction as general support that MLLMs handle short videos and is not used to justify Nar-KFC's central mechanism. Appendix A honestly concedes a training/test input-format misalignment, which is a limitation rather than a circular reduction. The most legitimate concern is protocol-level: the number of narratives (Fig. 4) and refinement window k (Table D.5) are selected on Video-MME, the same benchmark used for headline results, and Table 1 contains counterexamples (e.g., Qwen2-VL + Nar-KFC dropping on LVB and on Video-MME with subtitles). That is a soundness/data-snooping risk, not circularity, because the reported accuracies are not entailed by the method's construction. I therefore find no circular step under the defined criteria.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The method introduces no new physical entities. Its central claim rests on standard embedding and captioning models, plus several hand-set hyperparameters that were tuned on the evaluation benchmarks.

free parameters (6)
  • alpha = 1
    Balances query relevance and frame diversity in Eq (1); set by hand and not validated on a held-out set.
  • SVD rank r = N/4
    Number of retained singular values in the low-rank score matrix preprocessing in Alg 1; chosen as N/4 without a sensitivity study.
  • Downsample resolution = 128x128
    Fixed resolution of the downsampled score matrix in the greedy search; not ablated.
  • Refinement window k = 2
    Number of neighboring frames examined in the GS refinement step; Tab D.5 shows k=2 is best on Video-MME.
  • Number of narratives = 210
    Default caption count; Fig 4 shows accuracy rising with more captions and 210 was chosen as the default on Video-MME.
  • Captioner model and prompt = Qwen2-VL-2B, 15-word prompt
    Default captioner selected as a lightweight model; Fig 5 shows larger captioners give slightly higher accuracy, so the choice trades quality for cost.
assumptions (5)
  • domain assumption CLIP cosine similarity between frame and query is a valid query-relevance signal for keyframe selection.
    Section 3.1 uses SQR(i)=sim(f_i,q) as the first selection criterion; this assumes frozen CLIP embeddings capture what the question asks about.
  • domain assumption exp(-sim(f_i,f_j)) is a useful frame-diversity score.
    Section 3.1 defines SFD to avoid selecting near-duplicate frames; this assumes cosine distance in CLIP space reflects content similarity.
  • domain assumption Captions generated from non-keyframes preserve enough information to bridge temporal gaps between keyframes.
    Section 3.2 inserts non-keyframe captions into the MLLM prompt; Tab 5 shows pure narratives underperform frames, so the narrative stream adds value only if captions retain relevant details.
  • domain assumption MLLMs can process interleaved keyframes and narratives coherently despite no training on this format.
    Appendix A limitation (1) admits the training-testing misalignment; the entire Nar-KFC mechanism depends on this premise.
  • domain assumption Reported baselines from different evaluation toolkits are comparable to the authors' re-implemented numbers.
    Section 4 re-implements uniform baselines with VLMEvalKit, which 'may yield slight differences compared to other public toolkits'; Tab 1 mixes re-implemented and reported numbers.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Threading Keyframe with Narratives: MLLMs as Strong Long Video Comprehenders." pith.science (2026). https://pith.science/paper/K4JYYSFU

@misc{pith2026250524158,
  author       = {Pith},
  title        = {Pith review of: Threading Keyframe with Narratives: MLLMs as Strong Long Video Comprehenders},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K4JYYSFU}},
  note         = {Machine review of arXiv:2505.24158}
}
read the original abstract

Employing Multimodal Large Language Models (MLLMs) for long video understanding remains a challenging problem due to the dilemma between the substantial number of video frames (i.e., visual tokens) versus the limited context length of language models. Traditional uniform sampling often leads to selection of irrelevant content, while post-training MLLMs on thousands of frames imposes a substantial computational burden. In this paper, we propose threading keyframes with narratives (Nar-KFC), a plug-and-play module to facilitate effective and efficient long video perception. Nar-KFC generally involves two collaborative steps. First, we formulate the keyframe selection process as an integer quadratic programming problem, jointly optimizing query-relevance and frame-diversity. To avoid its computational complexity, a customized greedy search strategy is designed as an efficient alternative. Second, to mitigate the temporal discontinuity caused by sparse keyframe sampling, we further introduce interleaved textual narratives generated from non-keyframes using off-the-shelf captioners. These narratives are inserted between keyframes based on their true temporal order, forming a coherent and compact representation. Nar-KFC thus serves as a temporal- and content-aware compression strategy that complements visual and textual modalities. Experimental results on multiple long-video benchmarks demonstrate that Nar-KFC significantly improves the performance of popular MLLMs. Code will be made publicly available.

Figures

Figures reproduced from arXiv: 2505.24158 by the authors.

Figure 1
Figure 1. Illustration of keyframe capturing (KFC). SQR and SFD scores are computed via inner dot production. Uniform sampling is commonly used in short video under￾standing for consistent temporal structure. However, for long videos, it often misses important information with limited input. While recent works emphasize selecting query-relevant frames for long video QA, they tend to overlook the problem of narrow focus due to… view at source ↗
Figure 2
Figure 2. Illustration of Nar-KFC. We represent long videos by threading KFC-optimized keyframes with [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Accuracies (%) of uniform sampling, KFC, and Nar-KFC versus numbers of keyframes. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Impact of different captioners for gener [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Qualitative results. (left) Comparison of frames selected by uniform sampling, top-K sampling, and [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Training-free Uncertainty Guidance for Complex Visual Tasks with MLLMs

    cs.CV 2025-10 conditional novelty 6.0 of 10

    Selecting the visual input that minimizes an MLLM's output entropy (or maximizes its yes/no confidence) improves fine-grained visual search, long-video QA, and temporal grounding without any training.

Reference graph

Works this paper leans on

84 extracted references · 53 canonical work pages · cited by 1 Pith paper

  1. [1]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

    arXiv 2023

  2. [2]

    Flamingo: a visual language model for few-shot learning

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems (NeurIPS), 35:23716–23736, 2022

    2022

  3. [3]

    Vqa: Visual question answering

    Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE international conference on computer vision (ICCV), pages 2425–2433, 2015

    2015

  4. [4]

    Qwen-vl: A frontier large vision-language model with versatile abilities

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 1(2):3, 2023

    arXiv 2023

  5. [5]

    Solving mixed-integer quadratic programming problems with ibm-cplex: a progress report

    Christian Bliek1ú, Pierre Bonami, and Andrea Lodi. Solving mixed-integer quadratic programming problems with ibm-cplex: a progress report. In Proceedings of the twenty-sixth RAMP symposium, pages 16–17, 2014

    2014

  6. [6]

    On the opportunities and risks of foundation models

    Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021

    arXiv 2021

  7. [7]

    Language models are few-shot learners

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems (NeurIPS), 33:1877–1901, 2020

    1901

  8. [8]

    Hourvideo: 1-hour video-language understanding

    Keshigeyan Chandrasegaran, Agrim Gupta, Lea M Hadzic, Taran Kota, Jimming He, Cristóbal Eyzaguirre, Zane Durante, Manling Li, Jiajun Wu, and Fei-Fei Li. Hourvideo: 1-hour video-language understanding. Advances in Neural Information Processing Systems (NeurIPS), 37:53168–53197, 2024

    work page 2024

Show all 84 references
  1. [9]

    Sharegpt4video: Improving video understanding and generation with better captions

    Lin Chen, Xilin Wei, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Zhenyu Tang, Li Yuan, et al. Sharegpt4video: Improving video understanding and generation with better captions. Advances in Neural Information Processing Systems (NeurIPS), 37:19472...

    2024

  2. [10]

    Longvila: Scaling long-context visual language models for long videos

    Yukang Chen, Fuzhao Xue, Dacheng Li, Qinghao Hu, Ligeng Zhu, Xiuyu Li, Yunhao Fang, Haotian Tang, Shang Yang, Zhijian Liu, et al. Longvila: Scaling long-context visual language models for long videos. arXiv preprint arXiv:2408.10188, 2024

    2024 arXiv

  3. [11]

    Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vis...

    2024

  4. [12]

    Videollama 2: Advancing spatial-temporal modeling and audio understanding in video-llms

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

    2024 arXiv

  5. [13]

    Patch n’pack: Navit, a vision transformer for any aspect ratio and resolution

    Mostafa Dehghani, Basil Mustafa, Josip Djolonga, Jonathan Heek, Matthias Minderer, Mathilde Caron, Andreas Steiner, Joan Puigcerver, Robert Geirhos, Ibrahim M Alabdulmohsin, et al. Patch n’pack: Navit, a vision transformer for any aspect ratio and resolution. Advances in Neura...

    2023

  6. [14]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, G Heigold, S Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on...

    2020

  7. [15]

    Vlmevalkit: An open-source toolkit for evaluating large multi- modality models

    Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, Jiaqi Wang, et al. Vlmevalkit: An open-source toolkit for evaluating large multi- modality models. In Proceedings of the 32nd ACM International Conference on Multimedi...

    2024

  8. [16]

    Slowfast networks for video recognition

    Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV), pages 6202–6211, 2019

    2019

  9. [17]

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

    Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. In Proceedings of the IEEE/CVF Conference on...

    2025

  10. [18]

    The llama 3 herd of models

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al- Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024

    2024 arXiv

  11. [19]

    M-llm based video frame selection for efficient video understanding

    Kai Hu, Feng Gao, Xiaohan Nie, Peng Zhou, Son Tran, Tal Neiman, Lingyun Wang, Mubarak Shah, Raffay Hamid, Bing Yin, et al. M-llm based video frame selection for efficient video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition ...

    2025

  12. [20]

    Chat-univi: Unified visual representation empowers large language models with image and video understanding

    Peng Jin, Ryuichi Takanobu, Wancai Zhang, Xiaochun Cao, and Li Yuan. Chat-univi: Unified visual representation empowers large language models with image and video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages ...

    2024

  13. [21]

    Language repository for long video understanding

    Kumara Kahatapitiya, Kanchana Ranasinghe, Jongwoo Park, and Michael S Ryoo. Language repository for long video understanding. In Workshop on Video-Language Models@ NeurIPS 2024, 2024

    2024

  14. [22]

    An image grid can be worth a video: Zero-shot video question answering using a vlm

    Wonkyun Kim, Changin Choi, Wonseok Lee, and Wonjong Rhee. An image grid can be worth a video: Zero-shot video question answering using a vlm. IEEE Access, 2024

    2024

  15. [23]

    Lmms-eval: Accelerating the development of large multimoal models, March 2024

    Bo Li, Peiyuan Zhang, Kaichen Zhang, Fanyi Pu, Xinrun Du, Yuhao Dong, Haotian Liu, Yuanhan Zhang, Ge Zhang, Chunyuan Li, and Ziwei Liu. Lmms-eval: Accelerating the development of large multimoal models, March 2024

    2024

  16. [24]

    Llava-onevision: Easy visual task transfer

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

    2024 arXiv

  17. [25]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning (ICML), pages 19730–19742. PMLR, 2023

    2023

  18. [26]

    Videochat: Chat-centric video understanding

    KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Wenhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023

    2023 arXiv

  19. [27]

    Llama-vid: An image is worth 2 tokens in large language models

    Yanwei Li, Chengyao Wang, and Jiaya Jia. Llama-vid: An image is worth 2 tokens in large language models. In European Conference on Computer Vision (ECCV), pages 323–340. Springer, 2024

    2024

  20. [28]

    Video-llava: Learning united visual representation by alignment before projection

    Bin Lin, Yang Ye, Bin Zhu, Jiaxi Cui, Munan Ning, Peng Jin, and Li Yuan. Video-llava: Learning united visual representation by alignment before projection. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5971–5984, 2024

    2024

  21. [29]

    Vila: On pre- training for visual language models

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

    2024

  22. [30]

    Llava-next: Improved reasoning, ocr, and world knowledge, January 2024

    Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-next: Improved reasoning, ocr, and world knowledge, January 2024

    2024

  23. [31]

    Visual instruction tuning.Advances in neural information processing systems (NeurIPS), 36:34892–34916, 2023

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning.Advances in neural information processing systems (NeurIPS), 36:34892–34916, 2023. 11

    2023

  24. [32]

    Timecraft: Navigate weakly-supervised temporal grounded video question answering via bi-directional reasoning

    Huabin Liu, Xiao Ma, Cheng Zhong, Yang Zhang, and Weiyao Lin. Timecraft: Navigate weakly-supervised temporal grounded video question answering via bi-directional reasoning. In European Conference on Computer Vision (ECCV), pages 92–107. Springer, 2024

    2024

  25. [33]

    Lost in the middle: How language models use long contexts

    Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. Lost in the middle: How language models use long contexts. Transactions of the Association for Computational Linguistics (TACL), 12:157–173, 2024

    2024

  26. [34]

    St-llm: Large language models are effective temporal learners

    Ruyang Liu, Chen Li, Haoran Tang, Yixiao Ge, Ying Shan, and Ge Li. St-llm: Large language models are effective temporal learners. In European Conference on Computer Vision (ECCV), pages 1–18. Springer, 2024

    2024

  27. [35]

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

    Shuming Liu, Chen Zhao, Tianqi Xu, and Bernard Ghanem. Bolt: Boost large vision-language model without training for long-form video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

    2025

  28. [36]

    Drvideo: Document retrieval based long video understanding

    Ziyu Ma, Chenhui Gou, Hengcan Shi, Bin Sun, Shutao Li, Hamid Rezatofighi, and Jianfei Cai. Drvideo: Document retrieval based long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

    2025

  29. [37]

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

    Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. arXiv preprint arXiv:2306.05424, 2023

    2023 arXiv

  30. [38]

    Egoschema: A diagnostic benchmark for very long-form video language understanding

    Karttikeya Mangalam, Raiymbek Akshulakov, and Jitendra Malik. Egoschema: A diagnostic benchmark for very long-form video language understanding. Advances in Neural Information Processing Systems (NeurIPS), 36:46212–46244, 2023

    2023

  31. [39]

    Morevqa: Exploring modular reasoning models for video question answering

    Juhong Min, Shyamal Buch, Arsha Nagrani, Minsu Cho, and Cordelia Schmid. Morevqa: Exploring modular reasoning models for video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13235–13245, 2024

    2024

  32. [40]

    Branch-and-bound algorithms: A survey of recent advances in searching, branching, and pruning

    David R Morrison, Sheldon H Jacobson, Jason J Sauppe, and Edward C Sewell. Branch-and-bound algorithms: A survey of recent advances in searching, branching, and pruning. Discrete Optimization, 19:79–102, 2016

    2016

  33. [41]

    Chatgpt: Optimizing language models for dialogue, 2023

    OpenAI. Chatgpt: Optimizing language models for dialogue, 2023

    2023

  34. [42]

    Too many frames, not all useful: Efficient strategies for long-form video qa

    Jongwoo Park, Kanchana Ranasinghe, Kumara Kahatapitiya, Wonjeong Ryu, Donghyun Kim, and Michael S Ryoo. Too many frames, not all useful: Efficient strategies for long-form video qa. In Workshop on Video-Language Models@ NeurIPS 2024, 2024

    2024

  35. [43]

    Momentor: Advancing video large language model with fine-grained temporal reasoning

    Long Qian, Juncheng Li, Yu Wu, Yaobo Ye, Hao Fei, Tat-Seng Chua, Yueting Zhuang, and Siliang Tang. Momentor: Advancing video large language model with fine-grained temporal reasoning. In International Conference on Machine Learning (ICML), pages 41340–41356. PMLR, 2024

    2024

  36. [44]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning (IC...

    2021

  37. [45]

    The knapsack problem: a survey

    Harvey M Salkin and Cornelis A De Kluyver. The knapsack problem: a survey. Naval Research Logistics Quarterly, 22(1):127–144, 1975

    1975

  38. [46]

    Longvu: Spatiotemporal adaptive compression for long video-language understanding

    Xiaoqian Shen, Yunyang Xiong, Changsheng Zhao, Lemeng Wu, Jun Chen, Chenchen Zhu, Zechun Liu, Fanyi Xiao, Balakrishnan Varadarajan, Florian Bordes, et al. Longvu: Spatiotemporal adaptive compression for long video-language understanding. arXiv preprint arXiv:2410.17434, 2024

    arXiv

  39. [47]

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

    Yan Shu, Zheng Liu, Peitian Zhang, Minghao Qin, Junjie Zhou, Zhengyang Liang, Tiejun Huang, and Bo Zhao. Video-xl: Extra-long vision language model for hour-scale video understanding. arXiv preprint arXiv:2409.14485, 2024

    2024 arXiv

  40. [48]

    Two-stream convolutional networks for action recognition in videos

    Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. Advances in neural information processing systems (NeurIPS), 27, 2014

    2014

  41. [49]

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

    Enxin Song, Wenhao Chai, Guanhong Wang, Yucheng Zhang, Haoyang Zhou, Feiyang Wu, Haozhe Chi, Xun Guo, Tian Ye, Yanting Zhang, et al. Moviechat: From dense token to sparse memory for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Patt...

    2024

  42. [50]

    Mdp3: A training-free approach for list-wise frame selection in video-llms

    Hui Sun, Shiyin Lu, Huanyu Wang, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Ming Li. Mdp3: A training-free approach for list-wise frame selection in video-llms. arXiv preprint arXiv:2501.02885, 2025

    2025

  43. [51]

    Adaptive keyframe sampling for long video understanding

    Xi Tang, Jihao Qiu, Lingxi Xie, Yunjie Tian, Jianbin Jiao, and Qixiang Ye. Adaptive keyframe sampling for long video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

    2025

  44. [52]

    Gemma: Open models based on gemini research and technology

    Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, et al. Gemma: Open models based on gemini research and technology. arXiv preprint arXiv:2403.08295, 2024

    2024 arXiv

  45. [53]

    Cambrian-1: A fully open, vision- centric exploration of multimodal llms

    Peter Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Adithya Jairam Vedagiri IYER, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, et al. Cambrian-1: A fully open, vision- centric exploration of multimodal llms. Advances in Neural Information Process...

    2024

  46. [54]

    Llama 2: Open foundation and fine-tuned chat models

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023

    2023 arXiv

  47. [55]

    Attention is all you need

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

    2017

  48. [56]

    Show and tell: A neural image caption generator

    Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural image caption generator. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 3156–3164, 2015

    2015

  49. [57]

    Efficient large language models: A survey.Transactions on Machine Learning Research (TMLR), 2024

    Zhongwei Wan, Xin Wang, Che Liu, Samiul Alam, Yu Zheng, Jiachen Liu, Zhongnan Qu, Shen Yan, Yi Zhu, Quanlu Zhang, et al. Efficient large language models: A survey.Transactions on Machine Learning Research (TMLR), 2024

    2024

  50. [58]

    Weakly supervised gaussian contrastive grounding with large multimodal models for video question answering

    Haibo Wang, Chenghang Lai, Yixuan Sun, and Weifeng Ge. Weakly supervised gaussian contrastive grounding with large multimodal models for video question answering. In Proceedings of the 32nd ACM International Conference on Multimedia (MM), pages 5289–5298, 2024

    2024

  51. [59]

    Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution

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

    2024 arXiv

  52. [60]

    Retake: Reducing temporal and knowledge redundancy for long video understanding

    Xiao Wang, Qingyi Si, Jianlong Wu, Shiyu Zhu, Li Cao, and Liqiang Nie. Retake: Reducing temporal and knowledge redundancy for long video understanding. arXiv preprint arXiv:2412.20504, 2024

    2024 arXiv

  53. [61]

    Videoagent: Long-form video under- standing with large language model as agent

    Xiaohan Wang, Yuhui Zhang, Orr Zohar, and Serena Yeung-Levy. Videoagent: Long-form video under- standing with large language model as agent. In European Conference on Computer Vision (ECCV), pages 58–76. Springer, 2024

    2024

  54. [62]

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

    Ziyang Wang, Shoubin Yu, Elias Stengel-Eskin, Jaehong Yoon, Feng Cheng, Gedas Bertasius, and Mohit Bansal. Videotree: Adaptive tree-based video representation for llm reasoning on long videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition...

    2025

  55. [63]

    Dibs: Enhancing dense video captioning with unlabeled videos via pseudo boundary enrichment and online refinement

    Hao Wu, Huabin Liu, Yu Qiao, and Xiao Sun. Dibs: Enhancing dense video captioning with unlabeled videos via pseudo boundary enrichment and online refinement. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 18699–18708, 2024

    2024

  56. [64]

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

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

    2024

  57. [65]

    Next-qa: Next phase of question-answering to explaining temporal actions

    Junbin Xiao, Xindi Shang, Angela Yao, and Tat-Seng Chua. Next-qa: Next phase of question-answering to explaining temporal actions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 9777–9786, 2021

    2021

  58. [66]

    Can i trust your answer? visually grounded video question answering

    Junbin Xiao, Angela Yao, Yicong Li, and Tat-Seng Chua. Can i trust your answer? visually grounded video question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13204–13214, 2024. 13

    2024

  59. [67]

    Effective long-context scaling of foundation models

    Wenhan Xiong, Jingyu Liu, Igor Molybog, Hejia Zhang, Prajjwal Bhargava, Rui Hou, Louis Martin, Rashi Rungta, Karthik Abinav Sankararaman, Barlas Oguz, et al. Effective long-context scaling of foundation models. In Proceedings of the 2024 Conference of the North American Chapte...

    2024

  60. [68]

    Pllava: Parameter-free llava extension from images to videos for video dense captioning

    Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024

    2024 arXiv

  61. [69]

    Slowfast-llava: A strong training-free baseline for video large language models

    Mingze Xu, Mingfei Gao, Zhe Gan, Hong-You Chen, Zhengfeng Lai, Haiming Gang, Kai Kang, and Afshin Dehghan. Slowfast-llava: A strong training-free baseline for video large language models. arXiv preprint arXiv:2407.15841, 2024

    2024 arXiv

  62. [70]

    Zero-shot video question answering via frozen bidirectional language models

    Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. Advances in Neural Information Processing Systems (NeurIPS), 35:124–141, 2022

    2022

  63. [71]

    Vid2seq: Large-scale pretraining of a visual language model for dense video captioning

    Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision ...

    2023

  64. [72]

    Dense connector for mllms

    Huanjin Yao, Wenhao Wu, Taojiannan Yang, YuXin Song, Mengxi Zhang, Haocheng Feng, Yifan Sun, Zhiheng Li, Wanli Ouyang, and Jingdong Wang. Dense connector for mllms. Advances in Neural Information Processing Systems (NeurIPS), 37:33108–33140, 2024

    2024

  65. [73]

    Generative frame sampler for long video understanding

    Linli Yao, Haoning Wu, Kun Ouyang, Yuanxing Zhang, Caiming Xiong, Bei Chen, Xu Sun, and Junnan Li. Generative frame sampler for long video understanding. arXiv preprint arXiv:2503.09146, 2025

    2025 arXiv

  66. [74]

    Minicpm-v: A gpt-4v level mllm on your phone

    Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024

    2024 arXiv

  67. [75]

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

    Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. Advances in Neural Information Processing Systems (NeurIPS) , 36:76749–76771, 2023

    2023

  68. [76]

    Frame-voyager: Learning to query frames for video large language models

    Sicheng Yu, CHENGKAI JIN, Huanyu Wang, Zhenghao Chen, Sheng Jin, ZHONGRONG ZUO, XU XI- AOLEI, Zhenbang Sun, Bingni Zhang, Jiawei Wu, et al. Frame-voyager: Learning to query frames for video large language models. In The Thirteenth International Conference on Learning Represent...

    2024

  69. [77]

    Sigmoid loss for language image pre-training

    Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF international conference on computer vision (ICCV), pages 11975–11986, 2023

    2023

  70. [78]

    A simple llm framework for long-range video question-answering

    Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple llm framework for long-range video question-answering. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 21715–2...

    2024

  71. [79]

    Long context transfer from language to vision

    Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Chunyuan Li, and Ziwei Liu. Long context transfer from language to vision. arXiv preprint arXiv:2406.16852, 2024

    2024 arXiv

  72. [80]

    Llava-next: A strong zero-shot video understanding model, April 2024

    Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. Llava-next: A strong zero-shot video understanding model, April 2024

    2024

  73. [81]

    Video instruction tuning with synthetic data

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

    2024 arXiv

  74. [82]

    Mlvu: A comprehensive benchmark for multi-task long video understanding

    Junjie Zhou, Yan Shu, Bo Zhao, Boya Wu, Shitao Xiao, Xi Yang, Yongping Xiong, Bo Zhang, Tiejun Huang, and Zheng Liu. Mlvu: A comprehensive benchmark for multi-task long video understanding. arXiv preprint arXiv:2406.04264, 2024

    2024 arXiv

  75. [83]

    Minigpt-4: Enhancing vision-language understanding with advanced large language models

    Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023

    2023 arXiv

  76. [84]

    A. High heels

    Orr Zohar, Xiaohan Wang, Yann Dubois, Nikhil Mehta, Tong Xiao, Philippe Hansen-Estruch, Licheng Yu, Xiaofang Wang, Felix Juefei-Xu, Ning Zhang, et al. Apollo: An exploration of video understanding in large multimodal models. arXiv preprint arXiv:2412.10360, 2024. 14 Threading ...

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.