Beyond Frame Selection: Rethinking Long-Video Understanding with MLLMs

Shin'ichi Satoh; Ziling Huang

REVIEW 3 major objections 5 minor 13 references

Beyond Frame Selection: Rethinking Long-Video Understanding with MLLMs

T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Long-video QA works better by coordinating global and local evidence views than by picking one best set of frames.

desk verdict VideoRouter has a real observation about branch complementarity, but its headline 'same 64-frame budget' claim ignores that it uses 4.3x the MLLM calls per question. read the letter →

arxiv 2608.05592 v1 pith:NU2OVGHA submitted 2026-08-06 cs.CV

Ziling Huang , Shin'ichi Satoh This is my paper · ORCID

classification cs.CV

keywords long-videounderstandingmultimodallargelanguagemodelsframeselectiontemporalhierarchyverification-guidedroutingglobalandlocalevidenceviewsvideoquestionanswering

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

This paper argues that long-video question answering fails when treated as picking one best subset of frames, because questions need evidence at different time scales: some need broad storyline context, others need brief local details. The authors propose VideoRouter, a training-free method that organizes a video into a coarse-to-fine temporal hierarchy, derives a global evidence view from coarse segment representatives and a local evidence view by pruning the hierarchy toward question-relevant segments, and then verifies each branch's answer before routing to the final prediction. They show that a verification-guided router can coordinate these complementary views, and that on VideoMME it outscores the adaptive key-frame selection method AKS by 2.9 points with the same LLaVA-Video-7B backbone, while on LongVideoBench it beats uniform sampling by up to 8.2 points. The deeper claim is that branch complementarity, not single-branch strength, is the design objective: a locally weaker branch can be a better partner.

What carries the argument

The load-bearing mechanism is the verification-guided router operating on a temporal hierarchy. The hierarchy, built by recursive contiguous temporal clustering over CLIP frame features, exposes coarse story units at upper levels and fine clips at lower levels. From it, the global view takes representative center frames at depth drep, and the local view prunes segments whose max frame-question cosine similarity falls below alpha times the strongest segment, then rolls retained children up into parents. Each branch answers independently through an MLLM; a verifier MLLM then scores support in {0,1,2}; if one branch has strong support its answer is accepted, and otherwise a final MLLM pass runs over the union of both frame sets. This mechanism converts the complementarity trade-off into a routing decision rather than a frame-count optimization.

What would settle it

On a long-video QA dataset where the global and local branches retrieve exactly the same evidence, VideoRouter should perform no better than its stronger single branch, because the supposed complementarity is absent; a large improvement there would contradict the paper's coordination mechanism.

Watch

Extended reading notes

Core claim

The central claim, stated as the authors would state it, is that long-video understanding should be formulated as coordinating complementary evidence views, not as selecting a single optimal frame subset. VideoRouter instantiates this by building a question-agnostic temporal hierarchy via contiguous temporal clustering, constructing a global view from center frames of coarse segments and a local view from question-relevant pruned segments, and using an MLLM verifier that assigns a support score in {0,1,2} to each branch prediction, accepting a strongly supported branch and otherwise reasoning over the union of both evidence sets. The key observed phenomenon is that under a fixed global branch, stricter local pruning weakens the standalone local branch but raises the oracle upper bound, showing that a stronger local branch is not necessarily a better partner. A corollary is that the two benchmarks react oppositely--VideoMME favors global coverage, LongVideoBench favors local detail--yet the router improves on both.

Load-bearing premise

The load-bearing premise is that the verifier's support scores reliably indicate which branch's prediction is better supported by its selected frames; if the verifier is often wrong--the paper's own appendix puts its failure rate at 28.64% on VideoMME--routing adds noise instead of coordination.

Editorial extensions

If this is right

On VideoMME, VideoRouter improves over the AKS adaptive keyframe-selection method by 2.9 points under the same 64-frame budget with the LLaVA-Video-7B backbone.
On LongVideoBench, verification-guided routing beats uniform sampling by 3.41 to 8.23 points across four backbones.
The routing framework transfers across backbones, and verifier quality consistently matters: a stronger verifier improved routing by 0.15 to 0.85 points in the paper's controlled comparison.
Directly accepted samples are 13 to 22 points more accurate than fallback samples, indicating that verification successfully identifies cases where one branch already carries sufficient evidence.
The dataset-dependent preference--global coverage for VideoMME, local detail for LongVideoBench--shows that no single frame-selection strategy is universally best.

Reading between the lines

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

The reported 28.64% verifier failure rate on VideoMME points to headroom: strengthening or calibrating the verifier, rather than improving frame selection, may yield the next round of gains.
The visual-only temporal hierarchy could be extended to audio and subtitle signals, which the paper names as a limitation; richer story boundaries could make both branches more accurate.
A testable prediction follows from the complementarity claim: on any new long-video benchmark, varying local pruning strictness while holding the global view fixed should move the oracle upper bound opposite to the standalone local branch's accuracy.
The three-level support rubric is designed for multiple-choice QA; transferring it to open-ended generation would likely need a different verification protocol, a direction the paper leaves open.

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.

Desk Editor's Note

read the letter

VideoRouter is a decent training-free long-video QA method, and it has one genuinely interesting finding buried inside: the analysis around Figure 1 and Table 7 showing that a strictly pruned local branch is weaker on its own but produces a higher oracle upper bound when paired with a fixed global branch. That is a real observation about complementarity, and the verification-guided routing does seem to add something over next-token confidence and concatenation in their ablations. The hierarchy-based global/local construction is a reasonable new combination, even if each ingredient exists separately.

The soft spots are real, though, and the main one is compute. The paper's headline claim is that VideoRouter beats AKS under a "same 64-frame budget." That budget is per VQA call. VideoRouter makes 4.30 MLLM calls on average and consumes 28,921 visual tokens per question, versus 1 call and 16,384 tokens for Uniform-128; AKS under 64 frames is a single pass. The paper discloses these numbers in Appendix C.7 and frames the extra 3.08 seconds as moderate, but it never compares against a frame-selection method given the same total compute. If you give AKS or Q-Frame two or four times the frame budget and a verification pass, the +2.9 point VideoMME lead may well disappear. The conceptual argument is not falsified, but the empirical support is weaker than the "same budget" framing suggests.

Second, alpha=0.95 and d_rep=2 are selected after inspecting the same benchmark results they are used to report. That is tuning on the test set. The paper acknowledges this in the implementation details, but it should be flagged as a limitation. Third, the verifier failure rate of 28.64% on VideoMME is high; it does not kill the method because the router still beats simple baselines, but it means the routing decisions are noisy. Fourth, no code, no error bars.

None of this is load-bearing enough to reject outright. The paper is coherent, the experiments are honestly presented in the appendix, and the limitations section is thoughtful. It deserves a serious referee, but the referee should demand a compute-controlled comparison with equal total inference budget, a holdout-based hyperparameter selection, and error bars. I would bring it to a reading group to discuss evaluation methodology.

Referee Report

3 major / 5 minor

Summary. The manuscript proposes VideoRouter (VR), a training-free framework for multiple-choice long-video QA. It first organizes each video into a question-agnostic temporal hierarchy via contiguous clustering, then constructs two evidence views: a global view of segment-representative center frames and a local view obtained by pruning the hierarchy with a CLIP-based relevance threshold. The two views are answered independently by an MLLM; a verifier MLLM scores each answer's visual support (0/1/2), and routing either accepts the strongly supported branch or performs a third VQA pass on the union of both frame sets. Experiments on VideoMME and LongVideoBench across Qwen2.5-VL-3B, Qwen3-VL-4B, Qwen2-VL-7B, Qwen2.5-VL-7B, and LLaVA-Video-7B report consistent gains over uniform sampling and prior frame-selection and agent methods, with headline results of 64.4 on VideoMME and 63.6 on LongVideoBench for LLaVA-Video-7B under the claimed same-frame-budget protocol.

Significance. The central idea is interesting and the framework is clearly described. If confirmed under a properly budget-matched comparison, the global/local complementarity formulation and verification-guided routing would be a useful training-free addition to long-video MLLM literature. The paper has several strengths: it reports end-to-end token cost and latency (Table 10), includes a no-hierarchy control and lightweight ensemble baselines (Figure 3), analyzes verifier failure rates (Table 11), and candidly lists limitations in Appendix E. These elements allow the reader to see where the method could fail. However, the main empirical claim currently rests on a comparison that does not equalize inference compute, and the core routing mechanism has a nontrivial failure rate; both need to be addressed before the contribution can be taken at face value.

major comments (3)

[§4.2, Table 3; §C.7, Table 10] Table 3 is introduced as a comparison 'under the same 64-frame budget', but the budget is a per-inference frame cap, not a per-question cost. For LLaVA-Video-7B, VideoRouter runs two VQA branches, two verifier calls, and often a union fallback; Appendix C.7 (Table 10) reports for Qwen2.5-VL-7B on LongVideoBench an average of 4.30 MLLM calls and 28,921 visual tokens per question, versus 1 call and 16,384 tokens for Uniform-128. AKS and Q-Frame are single-call methods under the same per-call cap, so the +2.9-point VideoMME gain and the LVB gains in Table 3 may reflect a 3-4x increase in inference compute rather than the proposed global-local coordination. Please add a compute-controlled comparison (e.g., equal total visual tokens or equal number of MLLM calls, or AKS with iterative querying) and report the same cost columns for the Table 3 backbones.
[§4.1, Tables 6-7 and Figure 1] Hyperparameters that drive the final results, α=0.95 and d_rep=2, are selected using the same evaluation benchmarks. Section 4.1 states 'we choose α=0.95', and Tables 6-7 show that on VideoMME d_rep=4 gives Router 66.04 (vs 65.52 at d_rep=2) and α=0.90 gives 65.59; on LVB, d_rep=2 and α=0.95 are indeed best, but the VideoMME selection is not neutral. Since the headline 'outperforms frame selection' numbers use these chosen values, the comparison is not a clean holdout evaluation. Please either report a validation-split-based selection procedure, or report results across the full hyperparameter grid and justify the final choice independently of the test benchmarks.
[§3.4, §C.8, Appendix E] Verification-guided routing is the load-bearing component that decides direct acceptance vs. fallback, yet Appendix C.8 (Table 11) reports a verifier failure rate of 28.64% on VideoMME and 20.91% on LongVideoBench for cases where exactly one branch is correct. If the verifier picks the wrong branch on more than a quarter of decisive cases, the router can add noise rather than coordination; the paper's own limitations section (Appendix E) concedes that the coarse support levels 'may struggle on highly ambiguous cases'. Please provide a robustness analysis: report routing accuracy on the one-correct subset separately, compare verification routing against alternatives such as next-token confidence or agreement on that subset, and state how often the fallback rescues a wrong verifier decision. This would establish that the gains are not an artifact of a specific verifier.

minor comments (5)

[Abstract and §1] The phrase 'outperforms state-of-the-art frame selection methods by 2.9 points, respectively' is awkward; the LVB improvement over AKS (+1.5) is not mentioned, and the qualifier 'under the same 64-frame budget' is introduced only in Table 3. Please state both benchmark numbers and the exact comparison protocol in the abstract and introduction.
[§3.1] The text contains the typo 'Pseudu Code'; it should read 'Pseudo Code'.
[§C.3] The phrase 'tempoeal hierarchies' should be 'temporal hierarchies'.
[Appendix C.9] The sentence 'Luckily, table 6 shows, increasing frames in Local Branch can solve this problem partially' appears to refer to the wrong table; Table 6 is an ablation on global depth d_rep, not on local-branch frame count.
[Appendix C.6, Table 9] Table 9 reports average input frames for only two backbones, while the surrounding text says 'each VideoRouter branch uses substantially fewer frames'; extending the table to all reported backbones would make this claim checkable.

Circularity Check

1 steps flagged · score 2.0 of 10

Evaluation is largely external and not self-referential; one mild circularity from selecting α and d_rep on the same benchmarks used for headline results.

fitted input called prediction [Section 4.1 (Implementation Details); Appendix C.2/C.3, Tables 6 and 7]
"As shown in our introduction analysis, although a stricter local threshold slightly weakens the standalone local branch, it yields a higher oracle upper bound by making the local view more complementary to the global branch, so we choose α= 0.95."

The pruning ratio α (and similarly the global depth d_rep in Appendix C.2) is selected by inspecting router/oracle accuracies and oracle upper bounds computed on VideoMME and LongVideoBench, the same two benchmarks on which the headline gains (e.g., +2.9 over AKS) are reported. The final numbers are therefore not an independent out-of-sample evaluation of a fixed method: they are the accuracy at a parameter value chosen using the target benchmark labels. This is the 'fitted input called prediction' pattern in mild form, because the reported superiority is conditioned on benchmark-informed inputs rather than being forced by the method's equations.

full rationale

The paper's central comparison (VideoRouter vs. AKS/Q-Frame/uniform sampling) is run against external, fixed benchmarks and baselines, so the empirical claim is not self-referential in its core. There are no load-bearing self-citations and no imported uniqueness theorems; the temporal-hierarchy formulation is presented with its own construction and ablations. The main circularity-adjacent issue is that α=0.95 and d_rep=2 were chosen after inspecting the same VideoMME/LVB results that later appear as headline numbers, which mildly compromises the 'prediction' status of those numbers; this is a leakage/selection concern rather than an equation-level equivalence. The verifier sharing the QA backbone is a potential bias but is controlled against next-token confidence and agreement baselines, so it is not circular. The higher inference cost (4.30 MLLM calls, 1.77× visual tokens in Appendix C.7) is a fairness concern about the 'same 64-frame budget' framing, not a circularity. The Appendix C.8 verifier failure rate of 28.64% is a robustness risk, not circularity, because it is measured empirically against ground truth.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The method introduces no new physical or mathematical entities. Its free parameters are mostly benchmark-tuned hyperparameters and implementation choices. The key domain assumptions are unverified in the paper and one of them, the verifier reliability, is partially contradicted by the reported 28.64% failure rate.

free parameters (7)

local pruning ratio alpha = 0.95
Chosen in Section 4.1 based on oracle analysis in Figure 1 and ablations in Appendix C.3, on the same VideoMME/LVB benchmarks used for evaluation.
global depth d_rep = 2
Set as the default in Section 4.1, although Appendix C.2 shows d_rep=4 yields higher VideoMME accuracy (66.04 vs 65.52); the value is tuned on the evaluation benchmarks.
leaf length threshold L_leaf = 16
Recursion stop threshold in Section 3.1; no ablation or external justification.
local frames per segment = 16
Uniform sampling of 16 frames per retained segment in Section 3.3; no ablation reported.
frame sampling rate = 2.0 FPS
Sampling rate for hierarchy construction in Section 4.1; implementation choice.
max input frames cap = 128 (Qwen), 64 (LLaVA-Video)
Uniform subsampling caps in Section 4.1; chosen by GPU limits and model context.
number of coarse segments M
Not specified in the text; Figure 2 caption says the first layer typically contains around 32 coarse segments, but the method for choosing M is not given, leaving reimplementation ambiguous.

assumptions (5)

domain assumption CLIP visual features capture scene/story transitions and query relevance well enough for temporal clustering and local retrieval.
Invoked in Sections 3.1 and 3.3 where CLIP ViT-B/32 features are used both to build the hierarchy and to score segment relevance. If CLIP misses semantic boundaries, both views degrade.
domain assumption An MLLM verifier's support-level judgment is a trustworthy indicator of whether the selected frames support a candidate answer.
Core of verification-guided routing in Section 3.4; Appendix C.8 reports a 28.64% failure rate on VideoMME, so the assumption is partially contradicted by the paper's own data.
domain assumption Long-video evidence is multi-granular, with some questions depending on global event structure and others on short local details.
The premise of the whole design, stated in the Introduction and Section 3.1; not directly tested except through the benchmark results.
standard math Dynamic programming for contiguous segmentation is exact and runs on the sampled frame features.
Algorithm 1 in Appendix G uses standard interval DP with prefix sums to minimize SSE; this is a known technique.
domain assumption Frame-level max relevance r(s) identifies evidence-bearing moments in a segment.
Section 3.3 defines r(s) = max cosine similarity over frames in the segment; assumes the most relevant frame is representative of the segment's evidence.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Beyond Frame Selection: Rethinking Long-Video Understanding with MLLMs." pith.science (2026). https://pith.science/paper/NU2OVGHA

@misc{pith2026260805592,
  author       = {Pith},
  title        = {Pith review of: Beyond Frame Selection: Rethinking Long-Video Understanding with MLLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NU2OVGHA}},
  note         = {Machine review of arXiv:2608.05592}
}

read the original abstract

Multimodal Large Language Models (MLLMs) have achieved strong progress in video understanding, yet it remains challenging because the token limitation makes MLLMs difficult to capture temporally sparse evidence. Existing methods typically rely on uniform sampling, or frame selection, but these strategies usually optimize either broad temporal coverage or local relevance, making it difficult to preserve both global storyline context and fine-grained evidence. We propose VideoRouter(VR) that rethinks long-video understanding as coordinating complementary evidence views rather than selecting a single subset of frames. It first organizes each video into a question-agnostic temporal hierarchy, which partition the video into coarse-to-fine temporally coherent segments. In this hierarchy, upper-level nodes capture broad storyline context and event progression, while lower-level nodes preserve fine-grained local details and evidence-bearing moments. This naturally gives rise to two complementary views: a global view for coverage-oriented reasoning and a local view for detail-oriented evidence recovery. We further introduce a verification-guided router to determine which view is better supported by the selected evidence and select the final answer. We validate the effectiveness of the proposed design through extensive experiments, showing that the verification-guided router effectively coordinates global and local reasoning, and that, on VideoMME, our method outperforms state-of-the-art frame selection methods by 2.9 points, respectively, under the LLaVA-Video-7B backbone. We will release the code.

Figures

Figures reproduced from arXiv: 2608.05592 by the authors.

**Figure 1.** The best local branch is not always the best partner. Using the temporal hierarchy, we fix the global branch and vary the pruning threshold of the shared local branch. The x-axis denotes the strictness of local pruning, with α ∈ {0.85, 0.90, 0.95} from left to right. For the global branch, Global-U uniformly samples 4 frames from each segment at drep = 1, while Global-C uses the center frame of each segment at the s… view at source ↗

**Figure 2.** Overview of the proposed VideoRouter. The input video is first organized into a temporal hierarchy, where each s represents a segment containing multiple frames. For clarity, the illustration shows the full video being divided into two coarse segments at the first layer; in practice, however, the first layer typically contains around 32 coarse segments. Based on this hierarchy, we derive a global representative evid… view at source ↗

**Figure 3.** The global branch uniformly samples 64 frames because the corresponding hierarchy level [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: Qualitative visualization of global–local evidence selection and verification-guided routing. The gray arrow denotes the temporal axis. Local reasoning is more effective for detailcentric questions, while global reasoning is more reliable for questions involving broad…

**Figure 5.** Figure 5: Accuracy of direct acceptance and fallback reasoning. We report the accuracy of samples that are directly accepted by the verification-guided router (Accept) and those that require an additional union-based reasoning step (Fallback) on VideoMME and LVB. Across all back…

**Figure 6.** Figure 6: Coarse-to-fine construction of the temporal hierarchy. A video is first partitioned into [PITH_FULL_IMAGE:figures/full_fig_p020_6.png]

**Figure 7.** Figure 7: Prompt template for video multiple-choice QA. [PITH_FULL_IMAGE:figures/full_fig_p020_7.png]

**Figure 8.** Figure 8: Prompt template used by the verifier to assess the support level of a candidate answer given [PITH_FULL_IMAGE:figures/full_fig_p021_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

13 extracted references · 1 canonical work pages

[1]

Palm 2 technical report

Rohan Anil, Andrew M Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv:2305.10403,

arXiv
[4]

Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context.arXiv:2403.05530,

Gemini Team Google. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context.arXiv:2403.05530,

arXiv
[6]

Llava-onevision: Easy visual task transfer.arXiv:2408.03326,

10 Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer.arXiv:2408.03326,

arXiv
[7]

Commonsense video question answering through video-grounded entailment tree reasoning

Huabin Liu, Filip Ilievski, and Cees GM Snoek. Commonsense video question answering through video-grounded entailment tree reasoning. InCVPR, 2025a. Shuming Liu, Chen Zhao, Tianqi Xu, and Bernard Ghanem. Bolt: Boost large vision-language model without training for long-form video understanding. InProceedings of the IEEE/CVF Conference on Computer Vision a...

arXiv
[8]

Video-mtr: Reinforced multi-turn reasoning for long video understanding.arXiv:2508.20478,

Yuan Xie, Tianshui Chen, Zheng Ge, and Lionel Ni. Video-mtr: Reinforced multi-turn reasoning for long video understanding.arXiv:2508.20478,

arXiv
[9]

Qwen2 technical report

An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Ke-Yang Chen, Kexin Yang, Mei Li, Min Xue...

arXiv
[10]

mplug-owl: Modularization empowers large language models with multimodality.arXiv:2304.14178,

Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality.arXiv:2304.14178,

arXiv
[11]

Thinking with videos: Multimodal tool-augmented reinforcement learning for long video reasoning.arXiv:2508.04416, 2025a

Haoji Zhang, Xin Gu, Jiawen Li, Chixiang Ma, Sule Bai, Chubin Zhang, Bowen Zhang, Zhichao Zhou, Dongliang He, and Yansong Tang. Thinking with videos: Multimodal tool-augmented reinforcement learning for long video reasoning.arXiv:2508.04416, 2025a. Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Ch...

arXiv

Show all 13 references

[12]

Videolucy: Deep memory backtracking for long video understanding.arXiv:2510.12422,

Jialong Zuo, Yongtai Deng, Lingdong Kong, Jingkang Yang, Rui Jin, Yiwei Zhang, Nong Sang, Liang Pan, Ziwei Liu, and Changxin Gao. Videolucy: Deep memory backtracking for long video understanding.arXiv:2510.12422,
[13]

14 C.2 Ablation on the Global Representation Depth

12 APPENDIX A Model Checkpoints 14 B Datasets 14 C Experiments 14 C.1 Performance Breakdown by Verification Score Range . . . . . . . . . . . . . . . . 14 C.2 Ablation on the Global Representation Depth . . . . . . . . . . . . . . . . . . . . 15 C.3 Ablation on the Local Pruni...

2025
[2023]

Temporal chain of thought: Long-video understanding by thinking in frames.arXiv:2507.02001,

Anurag Arnab, Ahmet Iscen, Mathilde Caron, Alireza Fathi, and Cordelia Schmid. Temporal chain of thought: Long-video understanding by thinking in frames.arXiv:2507.02001,

arXiv
[2024]

Gpt-4o system card.arXiv:2410.21276,

Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Os- trow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card.arXiv:2410.21276,

arXiv
[2025]

Minigpt4-video: Advancing multimodal llms for video understanding with interleaved visual-textual tokens.arXiv:2404.03413, 2024a

Kirolos Ataallah, Xiaoqian Shen, Eslam Abdelrahman, Essam Sleiman, Deyao Zhu, Jian Ding, and Mohamed Elhoseiny. Minigpt4-video: Advancing multimodal llms for video understanding with interleaved visual-textual tokens.arXiv:2404.03413, 2024a. Kirolos Ataallah, Xiaoqian Shen, Es...

arXiv

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

Palm 2 technical report

Rohan Anil, Andrew M Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, et al. Palm 2 technical report. arXiv:2305.10403,

arXiv

[2] [4]

Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context.arXiv:2403.05530,

Gemini Team Google. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context.arXiv:2403.05530,

arXiv

[3] [6]

Llava-onevision: Easy visual task transfer.arXiv:2408.03326,

10 Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer.arXiv:2408.03326,

arXiv

[4] [7]

Commonsense video question answering through video-grounded entailment tree reasoning

Huabin Liu, Filip Ilievski, and Cees GM Snoek. Commonsense video question answering through video-grounded entailment tree reasoning. InCVPR, 2025a. Shuming Liu, Chen Zhao, Tianqi Xu, and Bernard Ghanem. Bolt: Boost large vision-language model without training for long-form video understanding. InProceedings of the IEEE/CVF Conference on Computer Vision a...

arXiv

[5] [8]

Video-mtr: Reinforced multi-turn reasoning for long video understanding.arXiv:2508.20478,

Yuan Xie, Tianshui Chen, Zheng Ge, and Lionel Ni. Video-mtr: Reinforced multi-turn reasoning for long video understanding.arXiv:2508.20478,

arXiv

[6] [9]

Qwen2 technical report

An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Ke-Yang Chen, Kexin Yang, Mei Li, Min Xue...

arXiv

[7] [10]

mplug-owl: Modularization empowers large language models with multimodality.arXiv:2304.14178,

Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality.arXiv:2304.14178,

arXiv

[8] [11]

Thinking with videos: Multimodal tool-augmented reinforcement learning for long video reasoning.arXiv:2508.04416, 2025a

Haoji Zhang, Xin Gu, Jiawen Li, Chixiang Ma, Sule Bai, Chubin Zhang, Bowen Zhang, Zhichao Zhou, Dongliang He, and Yansong Tang. Thinking with videos: Multimodal tool-augmented reinforcement learning for long video reasoning.arXiv:2508.04416, 2025a. Peiyuan Zhang, Kaichen Zhang, Bo Li, Guangtao Zeng, Jingkang Yang, Yuanhan Zhang, Ziyue Wang, Haoran Tan, Ch...

arXiv