Pith. sign in

REVIEW 3 major objections 5 minor 60 references

Treating video evidence as a set of intervals makes small models outperform 397B-parameter rivals.

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

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

T0 review · deepseek-v4-flash

2026-08-01 17:59 UTC pith:AF7JJ3EH

load-bearing objection TimeLens2's Wasserstein reward and staged data pipeline are real contributions; the headline SOTA rests on label quality that needs a human audit, but the paper is worth serious refereeing. the 3 major comments →

arxiv 2607.17423 v1 pith:AF7JJ3EH submitted 2026-07-19 cs.CV

TimeLens2: Generalist Video Temporal Grounding with Multimodal LLMs

Yuhan Zhu , Changlian Ma , Xiangyu Zeng , Xinhao Li , Zhiqiu Zhang , Songze Li , Jun Zhang , Tianxiang Jiang
show 7 more authors
Yuandong Yang Ziang Yan Zikang Wang Xinyu Chen Haoran Chen Shaowei Zhang Limin Wang
This is my paper
classification cs.CV
keywords video temporal groundingmultimodal large language modelsinterval-set supervisionWasserstein distancereinforcement learninglong-video understandingevidence verificationgeneralist grounding
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to establish that temporal grounding — saying *when* the supporting moments of a video happen — is a set-valued task, and that both supervision and optimization should treat it that way. The authors build a 93K-example training set by turning one brittle annotation pass into a cascade: caption-derived proposals, independent localization by two agents, cross-agent consensus, semantic verification, and local boundary refinement. They then add a reinforcement-learning reward based on the exact one-dimensional Wasserstein distance between uniform distributions over merged interval supports, giving dense, matching-free feedback for non-overlapping and multi-span predictions. On seven benchmarks representing long videos, recurring evidence, question-form queries, and egocentric video, the resulting 2B, 4B, and 8B models beat all size-matched baselines at 2B and reach state-of-the-art at 4B and 8B, surpassing far larger open-source models. If true, compact open models can make video-language outputs traceable to evidence without the scale of proprietary systems.

Core claim

The central claim is that the failure of video MLLMs at temporal grounding is a mismatch of representation: evidence is a variable-cardinality interval set, but labels are made by one global annotation decision and RL rewards either collapse all disjoint predictions to zero or require fragile one-to-one matching. TimeLens2 addresses the first mismatch with a staged pipeline that proposes coarse spans from timestamped captions, has two independent agents relocalize them from the video, keeps only instances with merged-support IoU above 0.9 and query-clip similarity above 0.5, then refines boundaries locally. It addresses the second with a reward that combines tIoU (precise overlap) with a tem

What carries the argument

The load-bearing object is the temporal Wasserstein reward R_TW. An interval set (possibly with several disjoint spans) is lifted to a uniform probability distribution over its merged support; the exact one-dimensional W1 distance between predicted and target distributions is computed in closed form as the integral of the absolute difference of their cumulative distribution functions, then scaled by target duration and exponentiated into a similarity. It supplies dense, matching-free geometric feedback: it ranks near misses above distant errors, guides disjoint predictions toward the target, and is invariant to how the same support is split into spans. The other half of the machinery is the

Load-bearing premise

The labels in TimeLens2-93K are correct: agreement between two independent annotators plus a text–video similarity check is assumed to certify all and only the true evidence intervals, even though the paper notes a shared semantic error between annotators cannot be ruled out.

What would settle it

Re-annotate a random subset of retained TimeLens2-93K instances with a third, stronger annotator (or human raters), asking for all supporting intervals; if a substantial fraction of instances contain repeated spans that both original agents missed, or if the same video content appears in evaluation under renamed IDs, the reported mIoU gains are inflated.

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

If this is right

  • Small (2B–8B) open models can reach state-of-the-art temporal grounding on long, multi-span, question-form, and egocentric videos, so grounding no longer requires a 100B+ model.
  • A single generalist model with one output interface can predict variable-cardinality interval evidence across domains, lengths, query forms, and viewpoints.
  • RL for temporal grounding benefits from a distance-aware reward: adding W1 rescues 75.8% of reinforcement-learning groups where tIoU gives all-zero reward, and recovers positive overlap for 21.9% of near-miss zero-overlap predictions.
  • Staged label curation is a quality lever: consensus, semantic verification, and boundary refinement improve mIoU while using one-eighth as many labels as the raw annotator output.
  • Output-format and instruction diversity during SFT teaches protocol-invariant interval semantics, improving transfer across benchmark prompt templates.

Where Pith is reading between the lines

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

  • If the staged verification recipe scales with stronger annotators, the same pipeline could produce larger corpora for spatiotemporal grounding, where evidence is localized in both time and space; the paper names this as a future direction but leaves it untested.
  • A testable extension: adding a third independently initialized annotator before the consensus gate should catch additional missed spans and improve recall-limited benchmarks, a prediction the current two-agent design does not test.
  • The Wasserstein reward's invariance to fragmentation implies it should also work for tasks with hierarchical or nested temporal evidence (e.g., action sequences with sub-events) without modifying the matching step.
  • Because the reported gains concentrate on zero-overlap cases, one might expect the reward's benefit to shrink on benchmarks where baselines already have high overlap; the paper's ablations do not measure this regime explicitly.

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 TimeLens2, a family of video multimodal LLMs for generalist temporal grounding, in which a model predicts a variable-cardinality set of evidence intervals. Two main contributions are claimed. First, TimeLens2-93K is a training corpus built from long YouTube videos through hierarchical captioning, segment-aware query synthesis, dual-agent localization, cross-agent consensus (IoUset > 0.9), semantic verification (query–clip cosine ≥ 0.5), and local boundary refinement. Second, a temporal Wasserstein reward is introduced, defined as the exact 1-D W1 distance between uniform distributions over merged interval supports, and is combined with tIoU and a parse penalty in GRPO. The authors report results on seven benchmarks (TimeLens-Bench re-annotations, VUE-TR/VUE-TR-V2, MomentSeeker text subset, Ego4D-NLQ), claiming that TimeLens2-2B outperforms all size-matched baselines and that the 4B/8B variants surpass open models up to 397B parameters, with consistency across video lengths, query forms, and viewpoints. The supplementary material includes full threshold breakdowns, a train–test overlap audit, and a limitations section.

Significance. If the results hold, this is a valuable contribution: the paper gives a clean, matching-free, fragmentation-invariant reward for multi-interval temporal grounding, and it demonstrates a plausible staged-pipeline recipe for constructing long-video grounding supervision. The exact W1 CDF form is mathematically sound, and the reward ablations (Tables 6–9) are internally consistent and mechanistically informative. The paper ships code, model weights, and a dataset, and its data-scaling and context-length ablations are useful for the community. However, the headline claims rest on the correctness of the auto-generated TimeLens2-93K labels and on the fairness of the benchmark comparisons; both need strengthening before the results can be taken at face value.

major comments (3)
  1. [Section 3.1; Supplement A.5, A.6] The load-bearing assumption is that TimeLens2-93K labels are approximately correct. The verification pipeline retains instances with IoUset(Qwen3-VL-30B-A3B, TimeLens-8B) > 0.9 and normalized query–clip cosine ≥ 0.5, but neither threshold is calibrated against human judgment. The paper itself states that agreement 'cannot rule out a shared semantic error' (Section 3.1) and, in A.6, that corpus quality is bounded by the annotating models. Supplement A.5 additionally admits that for MomentSeeker, renamed identifiers prevent ruling out content duplicates. Systematic shared errors in the surviving labels would inflate every reported mIoU gain, especially the multi-span and long-video gains. The authors should provide a human audit on a stratified sample of TimeLens2-93K, reporting per-stratum precision/recall of interval sets, agreement with the pipeline labels, and the operating point of th
  2. [Section 4.2; Tables 1, 6, 7] All headline numbers appear to be single-run, self-reported evaluations. No standard errors, confidence intervals, or multiple seeds are reported, even though several differences are small: the Wasserstein reward improves average mIoU over tIoU by only 0.7 points (Table 6), and over matched NGIoU by 0.6 points (Table 7); TimeLens2-8B trails Gemini 2.5 Pro by 0.2 mIoU on QVHighlights in Table 1. The claims that Wasserstein is superior to NGIoU, or that TimeLens2 'outperforms all size-matched baselines on every benchmark,' require at least bootstrap intervals or paired significance tests over the benchmark queries. Additionally, baselines are evaluated in-house with model-specific input overrides (Section A.4); the reproducibility of those numbers would be improved by reporting the exact code/config versions and any deviations from each baseline's native protocol.
  3. [Section 4.1; Supplement A.5] The training mixture is TimeLens2-93K, TimeLens-100K, and the Ego4D-NLQ training split. The train–test overlap audit in A.5 checks recoverable YouTube IDs and reports zero overlap, but it does not explicitly address whether TimeLens-100K contains videos from, or near-duplicates of, the evaluation benchmarks, nor does it resolve renamed MomentSeeker content. It also evaluates on TimeLens-Bench, which is the authors' own re-annotation of Charades-STA, ActivityNet, and QVHighlights; using a benchmark re-annotated by the same group to validate the model can narrow the test distribution toward their annotation protocol. The paper should provide a complete overlap table for every training component against every evaluation subset, including fuzzy-duplicate detection, and state whether TimeLens2-93K has a held-out split.
minor comments (5)
  1. [Section 3.2, Eq. (8)] The ε in RTW = exp(−W/(|merge(Y)|+ε)) is not given a value. Please state the numerical value used in experiments and whether results are sensitive to it.
  2. [Figure 2] The pipeline diagram is visually dense; many labels are too small to read at print size. Consider splitting into subfigures or enlarging the caption text.
  3. [Section 4.1] Table 12's FPS and frames_limit values suggest that for very long videos the effective sampling rate is lower than the nominal FPS. Please clarify the effective sampling rule in the text, since it affects comparability across benchmarks.
  4. [Tables 6/7] The tables label the reward as R = R_temporal − 1 invalid. Please clarify whether the tIoU baseline also includes the fixed −1 invalid penalty; the text says this penalty is fixed in all RL experiments, but the table notation could be read as excluding it from the baseline.
  5. [Abstract / Section 4.2] The abstract states TimeLens2-2B 'outperforms all size-matched baselines on every benchmark,' but Table 1 compares only three 2B baselines. This is correct but should be phrased as 'all evaluated size-matched baselines' to avoid overgeneralization.

Circularity Check

0 steps flagged

No significant circularity; the claimed contributions are empirically grounded and not reduced to their inputs by construction.

full rationale

The paper's derivation chain does not contain a step in which a prediction or claimed result is equivalent by construction to a fitted input or to a self-citation. The data-construction gates (IoUset(Ŷ_Q,Ŷ_T)>0.9 and normalized cosine similarity ≥0.5) are label-filtering thresholds; they do not define the evaluation metric or the model output, and they are not fitted to benchmark scores. The temporal Wasserstein reward R_TW = exp(−W/(|merge(Y)|+ε)) is a fixed mathematical objective with no parameters fitted to the evaluation benchmarks, and its contribution is isolated by ablations against R_tIoU and NGIoU on held-out benchmarks. The reported gains are empirical measurements, not consequences of defining the target in terms of the prediction. The use of TimeLens-8B, TimeLens-100K, and TimeLens-Bench from prior work by overlapping authors is a practical reuse of resources, but the load-bearing argument does not reduce to those citations: the ablations show the new TimeLens2-93K pipeline and the Wasserstein reward each improve performance independently, including on external benchmarks (VUE-TR, VUE-TR-V2, MomentSeeker, Ego4D-NLQ). The paper's own limitations—that label quality is bounded by annotator models, that cross-agent agreement cannot rule out shared semantic error, and that renamed-content overlap cannot be excluded—are honest data-quality caveats, not circular reductions by definition or by self-citation chain.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central derivation depends on one standard-math result (1-D Wasserstein CDF formula), two domain-representation choices (uniform temporal mass; consensus-as-truth), and the reliability of benchmark labels and no-leakage assumptions. The main hand-set free parameters are the label-curation thresholds and refinement window; no new physical or conceptual entities are introduced.

free parameters (4)
  • Temporal consensus threshold = IoUset > 0.9
    Section 3.1: instances are retained only when the two annotators' merged supports agree above 0.9. Hand-chosen; no sensitivity analysis is reported, and this threshold controls which training labels survive.
  • Semantic verification threshold = cosine >= 0.5
    Section 3.1: a normalized Qwen3-VL-Embedding text-video cosine similarity of at least 0.5 is required for retained labels. Hand-chosen; no calibration or ablation is reported.
  • Boundary refinement window = +/- 3 seconds
    Section 3.1: refinement is restricted to a +/-3-second neighborhood around each verified boundary. This hand-set window defines the maximum possible label shift.
  • Gap-merge threshold = 1 second
    Section 3.1: adjacent spans separated by at most one second are merged as boundary jitter. This is a hand-set label-formatting choice that affects multi-span supervision.
axioms (5)
  • standard math Exact 1-Wasserstein distance over one-dimensional distributions is equal to the integral of the absolute difference of their CDFs.
    Invoked in Section 3.2 in the equation for W(Ŷ,Y); this is standard optimal-transport theory and does not depend on the paper's data.
  • domain assumption Temporal evidence is best represented as uniform probability mass over the merged interval support, so that equivalent fragmentation has identical reward.
    Section 3.2 and Figure 3: the paper chooses uniform distributions over merged supports rather than endpoint atoms or Gaussians; this modeling choice is exactly what makes the reward fragmentation-invariant.
  • domain assumption Agreement between two independently trained localizers plus an embedding-similarity gate is sufficient evidence that an interval set is a correct label.
    Section 3.1: the pipeline keeps labels only when IoUset>0.9 and cosine>=0.5. The paper itself concedes that agreement cannot rule out shared semantic errors, making this a load-bearing but unproven assumption.
  • domain assumption The benchmark ground-truth interval sets are complete and correct for each query.
    All evaluation numbers in Tables 1 and 11 assume that the benchmark labels define the true evidence. If any benchmark has missing or noisy intervals, the mIoU comparisons are distorted.
  • domain assumption No training video shares content with evaluation videos beyond recoverable exact ID matches.
    Supplementary A.5 finds zero exact ID overlap but explicitly states it cannot exclude content duplicates concealed by renaming. The integrity of the SOTA comparison depends on this unverifiable premise.

pith-pipeline@v1.3.0-alltime-deepseek · 27956 in / 10620 out tokens · 113148 ms · 2026-08-01T17:59:03.473692+00:00 · methodology

0 comments
read the original abstract

Video multimodal large language models (MLLMs) can describe what happens in a video, but rarely identify when the supporting evidence occurs. We study generalist video temporal grounding, in which one model predicts a variable-cardinality set of evidence intervals across video lengths, domains, query forms, and viewpoints. Existing training strategies are misaligned with this set-valued task: long-video labels often rely on brittle one-pass annotation, while reinforcement-learning rewards either fail to distinguish non-overlapping predictions or require fragile segment matching. TimeLens2 treats temporal evidence as an interval set throughout supervision and optimization. TimeLens2-93K constructs reliable multi-span supervision through caption-derived proposals, independent localization, cross-agent consensus, semantic verification, and boundary refinement. Our temporal Wasserstein reward computes exact one-dimensional \(W_1\) between uniform distributions over merged interval supports, providing dense, matching-free feedback under unequal cardinalities and equivalent fragmentation; temporal IoU complements it with precise-overlap feedback. Across seven benchmarks, TimeLens2-2B outperforms all size-matched baselines on every benchmark, while the 4B and 8B variants achieve state-of-the-art performance, surpassing open-source models with up to 397B parameters. The 2B, 4B, and 8B variants improve over their Qwen3-VL backbones by 14.2, 13.0, and 18.1 mIoU points, respectively.

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

60 extracted references · 1 canonical work pages

  1. [1]

    Llava-onevision-2: Towards next-generation perceptual intelligence.arXiv preprint arXiv:2605.25979, 2026

    Xiang An, Yin Xie, Feilong Tang, Yunyao Yan, Huajie Tan, Didi Zhu, Changrui Chen, Xiuwei Zhao, Bin Qin, Kaicheng Yang, et al. Llava-onevision-2: Towards next-generation perceptual intelligence.arXiv preprint arXiv:2605.25979, 2026

    Pith/arXiv arXiv 2026

  2. [2]

    Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025

    Pith/arXiv arXiv 2025

  3. [3]

    Datasets and recipes for video temporal grounding via reinforcement learning, 2025

    Ruizhe Chen, Zhiting Fan, Tianze Luo, Heqing Zou, Zhaopeng Feng, Guiyang Xie, Hansheng Zhang, Zhuochen Wang, Zuozhu Liu, and Huaijian Zhang. Datasets and recipes for video temporal grounding via reinforcement learning, 2025. URL https://arxiv.org/abs/2507 .18100

    2025

  4. [4]

    Molmo2: Open weights and data for vision-language models with video understanding and grounding

    Christopher Clark, Jieyu Zhang, Zixian Ma, Jae Sung Park, Rohun Tripathi, Sangho Lee, Mo- hammadreza Salehi, Jason Ren, Chris Dongjoo Kim, Yinuo Yang, et al. Molmo2: Open weights and data for vision-language models with video understanding and grounding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 28652–28668, 2026

    2026

  5. [5]

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507 .06261, 2025

    Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507 .06261, 2025

    2025

  6. [6]

    Videotg-r1: Boosting video temporal grounding via curriculum reinforcement learning on reflected boundary annotations, 2025

    Lu Dong, Haiyu Zhang, Han Lin, Ziang Yan, Xiangyu Zeng, Hongjie Zhang, Yifei Huang, Yi Wang, Zhen-Hua Ling, Limin Wang, and Yali Wang. Videotg-r1: Boosting video temporal grounding via curriculum reinforcement learning on reflected boundary annotations, 2025. URL https: //arxiv.org/abs/2510.23397

    arXiv 2025

  7. [7]

    Vlmevalkit: An open-source toolkit for evaluating large multi-modality models, 2026

    Haodong Duan, Xinyu Fang, Junming Yang, Xiangyu Zhao, Zerun Ma, Yuxuan Qiao, Mo Li, Tian- hao Liang, Lin Zhu, Amit Agarwal, Xiaozhe Li, Shengyuan Ding, Jiazi Bu, Ziyu Liu, Zhangyang Qi, Yifei Li, Yuhang Zang, Zhe Chen, Lin Chen, Yuan Liu, Yubo Ma, Hailong Sun, Yifan Zhang, Shiyin Lu, Tack Hwa Wong, Weiyun Wang, Peiheng Zhou, Chaoyou Fu, Junbo Cui, Jixuan ...

    2026

  8. [8]

    Tall: Temporal activity localization via language query

    Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. InProceedings of the IEEE international conference on computer vision, pp. 5267–5275, 2017

    2017

  9. [9]

    Gemini 3: News and announcements

    Google. Gemini 3: News and announcements. https://blog.google/products-and-platforms/ products/gemini/gemini-3-collection/, 2025

    2025

  10. [10]

    Gemini 3.1 Pro model card

    Google DeepMind. Gemini 3.1 Pro model card. https://deepmind.google/models/model-cards/ gemini-3-1-pro, 2026

    2026

  11. [11]

    Ego4d: Around the world in 3,000 hours of egocentric video

    Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 18995–19012, 2022

    2022

  12. [12]

    Localizing moments in video with natural language, 2017

    Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language, 2017 . URL https://arxiv.org/abs/1708.01641

    Pith/arXiv arXiv 2017

  13. [13]

    Vtimellm: Empower llm to grasp video moments

    Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 14271–14280, 2024

    2024

  14. [14]

    Lita: Language instructed temporal-localization assistant, 2024

    De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant, 2024. URL https://arxiv.org/abs/2403.19046. 13 TimeLens2 REFERENCES

    Pith/arXiv arXiv 2024

  15. [15]

    Dense-captioning events in videos

    Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. InProceedings of the IEEE international conference on computer vision, pp. 706–715, 2017

    2017

  16. [16]

    Berg, and Mohit Bansal

    Jie Lei, Licheng Yu, Tamara L. Berg, and Mohit Bansal. Tvr: A large-scale dataset for video- subtitle moment retrieval, 2020. URL https://arxiv.org/abs/2001.09099

    Pith/arXiv arXiv 2020

  17. [17]

    Detecting moments and highlights in videos via natural language queries.Advances in Neural Information Processing Systems, 34:11846–11858, 2021

    Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries.Advances in Neural Information Processing Systems, 34:11846–11858, 2021

    2021

  18. [18]

    Re- inforcement learning tuning for videollms: Reward design and data efficiency.arXiv preprint arXiv:2506.01908, 2025

    Hongyu Li, Songhao Han, Yue Liao, Junfeng Luo, Jialin Gao, Shuicheng Yan, and Si Liu. Re- inforcement learning tuning for videollms: Reward design and data efficiency.arXiv preprint arXiv:2506.01908, 2025

    Pith/arXiv arXiv 2025

  19. [19]

    Video-opd: Efficient post-training of multimodal large language models for temporal video grounding via on-policy distillation, 2026

    Jiaze Li, Hao Yin, Haoran Xu, Boshen Xu, Wenhui Tan, Zewen He, Jianzhong Ju, Zhenbo Luo, and Jian Luan. Video-opd: Efficient post-training of multimodal large language models for temporal video grounding via on-policy distillation, 2026. URL https://arxiv.org/abs/2602.02994

    Pith/arXiv arXiv 2026

  20. [20]

    Qwen3-vl-embedding and qwen3-vl-reranker: A unified framework for state-of-the-art multimodal retrieval and ranking,

    Mingxin Li, Yanzhao Zhang, Dingkun Long, Keqin Chen, Sibo Song, Shuai Bai, Zhibo Yang, Pengjun Xie, An Yang, Dayiheng Liu, Jingren Zhou, and Junyang Lin. Qwen3-vl-embedding and qwen3-vl-reranker: A unified framework for state-of-the-art multimodal retrieval and ranking,

  21. [21]

    Videochat-flash: Hierarchical compression for long-context video modeling.arXiv preprint arXiv:2501.00574, 2024

    Xinhao Li, Yi Wang, Jiashuo Yu, Xiangyu Zeng, Yuhan Zhu, Haian Huang, Jianfei Gao, Kunchang Li, Yinan He, Chenting Wang, et al. Videochat-flash: Hierarchical compression for long-context video modeling.arXiv preprint arXiv:2501.00574, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Videochat-r1: Enhancing spatio-temporal perception via reinforce- ment fine-tuning, 2025

    Xinhao Li, Ziang Yan, Desen Meng, Lu Dong, Xiangyu Zeng, Yinan He, Yali Wang, Yu Qiao, Yi Wang, and Limin Wang. Videochat-r1: Enhancing spatio-temporal perception via reinforce- ment fine-tuning, 2025. URL https://arxiv.org/abs/2504.06958

    Pith/arXiv arXiv 2025

  23. [23]

    Videochat3: Fully open video mllm for efficient and generalist video understanding, 2026

    Xinhao Li, Yuhan Zhu, Xiangyu Zeng, Yuhao Dong, Haoning Wu, Zhiqiu Zhang, Yuandong Yang, Changlian Ma, Qingyu Zhang, Yansong Shi, Xinyu Chen, Haoran Chen, Zizheng Huang, Jun Zhang, Kun Ouyang, Lin Sui, Ziang Yan, Yicheng Xu, Chenting Wang, Yinan He, Hongjie Zhang, Yi Wang, Yu Qiao, Yali Wang, Ziwei Liu, Kai Chen, and Limin Wang. Videochat3: Fully open vid...

    2026

  24. [24]

    Reasoning guided embeddings: Leveraging mllm reasoning for improved multimodal retrieval,

    Chunxu Liu, Jiyuan Yang, Ruopeng Gao, Yuhan Zhu, Feng Zhu, Rui Zhao, and Limin Wang. Reasoning guided embeddings: Leveraging mllm reasoning for improved multimodal retrieval,

  25. [25]

    Museg: Reinforcing video temporal understanding via timestamp-aware multi-segment grounding

    Fuwen Luo, Shengfeng Lou, Chi Chen, Ziyue Wang, Chenliang Li, Weizhou Shen, Jiyue Guo, Peng Li, Ming Yan, Ji Zhang, Fei Huang, and Yang Liu. Museg: Reinforcing video temporal understanding via timestamp-aware multi-segment grounding. InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 3554...

    work page doi:10.18653/v1/2026.acl-long.1644 2026

  26. [26]

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

    Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad 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 (Volume 1: Long Papers), pp. 12585–12602, 2024

    2024

  27. [27]

    Howto100m: Learning a text-video embedding by watching hundred million narrated video clips, 2019

    Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips, 2019 . URL https://arxiv.org/abs/1906.03327

    Pith/arXiv arXiv 2019

  28. [28]

    Marlin-2B: A tiny vlm to extract structured information from videos

    NemoStation. Marlin-2B: A tiny vlm to extract structured information from videos. Hugging Face model repository, 2026. URL https://huggingface.co/NemoStation/Marlin-2B. Accessed July 18, 2026. 14 TimeLens2 REFERENCES

    2026

  29. [29]

    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,

  30. [30]

    Grounding action descriptions in videos.Transactions of the Association for Computational Linguistics, 1:25–36, 2013

    Michaela Regneri, Marcus Rohrbach, Dominikus Wetzel, Stefan Thater, Bernt Schiele, and Manfred Pinkal. Grounding action descriptions in videos.Transactions of the Association for Computational Linguistics, 1:25–36, 2013

    2013

  31. [31]

    Timechat: A time-sensitive multi- modal large language model for long video understanding.ArXiv, abs/2312.02051, 2023

    Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multi- modal large language model for long video understanding.ArXiv, abs/2312.02051, 2023

    Pith/arXiv arXiv 2023

  32. [32]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y . K. Li, Y . Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models, 2024. URL https://arxiv.org/abs/2402.03300

    Pith/arXiv arXiv 2024

  33. [33]

    Openai gpt-5 system card.arXiv preprint arXiv:2601.03267, 2025

    Aaditya Singh, Adam Fry, Adam Perelman, Adam Tart, Adi Ganesh, Ahmed El-Kishky, Aidan McLaughlin, Aiden Low, AJ Ostrow, Akhila Ananthram, et al. Openai gpt-5 system card.arXiv preprint arXiv:2601.03267, 2025

    Pith/arXiv arXiv 2025

  34. [34]

    Mad: A scalable dataset for language grounding in videos from movie audio descriptions

    Mattia Soldan, Alejandro Pardo, Juan León Alcázar, Fabian Caba, Chen Zhao, Silvio Giancola, and Bernard Ghanem. Mad: A scalable dataset for language grounding in videos from movie audio descriptions. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 5026–5035, 2022

    2022

  35. [35]

    Coin: A large-scale dataset for comprehensive instructional video analysis, 2019

    Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis, 2019 . URL https://arxiv.org/abs/1903.02874

    Pith/arXiv arXiv 2019

  36. [36]

    Kimi Team, Tongtong Bai, Yifan Bai, Yiping Bao, SH Cai, Yuan Cao, Y Charles, HS Che, Cheng Chen, Guanduo Chen, et al. Kimi k2. 5: Visual agentic intelligence.arXiv preprint arXiv:2602.02276, 2026

    Pith/arXiv arXiv 2026

  37. [37]

    Qwen3.5: Accelerating productivity with native multimodal agents, February 2026

    Qwen Team. Qwen3.5: Accelerating productivity with native multimodal agents, February 2026. URL https://qwen.ai/blog?id=qwen3.5

    2026

  38. [38]

    Vidi: Large multimodal models for video understanding and editing.arXiv preprint arXiv:2504.15681, 2025

    Vidi Team, Celong Liu, Chia-Wen Kuo, Dawei Du, Fan Chen, Guang Chen, Jiamin Yuan, Lingxi Zhang, Lu Guo, Lusha Li, et al. Vidi: Large multimodal models for video understanding and editing.arXiv preprint arXiv:2504.15681, 2025

    Pith/arXiv arXiv 2025

  39. [39]

    Vidi2: Large multimodal models for video understanding and creation.arXiv preprint arXiv:2511.19529, 2025

    Vidi Team, Celong Liu, Chia-Wen Kuo, Chuang Huang, Dawei Du, Fan Chen, Guang Chen, Haoji Zhang, Haojun Zhao, Lingxi Zhang, et al. Vidi2: Large multimodal models for video understanding and creation.arXiv preprint arXiv:2511.19529, 2025

    arXiv 2025

  40. [40]

    Internvideo-next: Towards general video foundation models without video- text supervision, 2026

    Chenting Wang, Yuhan Zhu, Yicheng Xu, Jiange Yang, Lang Lin, Ziang Yan, Yali Wang, Yi Wang, and Limin Wang. Internvideo-next: Towards general video foundation models without video- text supervision, 2026. URL https://arxiv.org/abs/2512.01342

    arXiv 2026

  41. [41]

    Grounded-videollm: Sharpening fine-grained temporal grounding in video large language models, 2025

    Haibo Wang, Zhiyang Xu, Yu Cheng, Shizhe Diao, Yufan Zhou, Yixin Cao, Qifan Wang, Weifeng Ge, and Lifu Huang. Grounded-videollm: Sharpening fine-grained temporal grounding in video large language models, 2025. URL https://arxiv.org/abs/2410.03290

    Pith/arXiv arXiv 2025

  42. [42]

    A normalized gaussian wasserstein distance for tiny object detection.arXiv preprint arXiv:2110.13389, 2021

    Jinwang Wang, Chang Xu, Wen Yang, and Lei Yu. A normalized gaussian wasserstein distance for tiny object detection.arXiv preprint arXiv:2110.13389, 2021

    Pith/arXiv arXiv 2021

  43. [43]

    Internvl3

    Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Ye, Jie Shao, et al. Internvl3. 5: Advancing open-source multimodal models in versatility, reasoning, and efficiency.arXiv preprint arXiv:2508.18265, 2025

    Pith/arXiv arXiv 2025

  44. [44]

    Time-r1: Post-training large vision language model for temporal video grounding, 2025

    Ye Wang, Ziheng Wang, Boshen Xu, Yang Du, Kejun Lin, Zihan Xiao, Zihao Yue, Jianzhong Ju, Liang Zhang, Dingyi Yang, Xiangnan Fang, Zewen He, Zhenbo Luo, Wenxuan Wang, Junqi Lin, Jian Luan, and Qin Jin. Time-r1: Post-training large vision language model for temporal video grounding, 2025. URL https://arxiv.org/abs/2503.13377. 15 TimeLens2 REFERENCES

    Pith/arXiv arXiv 2025

  45. [45]

    Tempr1: Improving temporal understanding of mllms via temporal-aware multi-task reinforcement learning, 2026

    Tao Wu, Li Yang, Gen Zhan, Yabin Zhang, Yiting Liao, Junlin Li, Deliang Fu, Li Zhang, and Limin Wang. Tempr1: Improving temporal understanding of mllms via temporal-aware multi-task reinforcement learning, 2026. URL https://arxiv.org/abs/2512.03963

    Pith/arXiv arXiv 2026

  46. [46]

    Mimo-vl technical report, 2025

    LLM-Core-Team Xiaomi. Mimo-vl technical report, 2025. URL https://arxiv.org/abs/2506.03569

    Pith/arXiv arXiv 2025

  47. [47]

    Internvideo3: Agentify foundation models with multimodal contextual reasoning, 2026

    Ziang Yan, Sheng Xia, Jiashuo Yu, Yue Wu, Tianxiang Jiang, Songze Li, Kanghui Tian, Yicheng Xu, Yinan He, Kai Chen, Limin Wang, Yu Qiao, and Yi Wang. Internvideo3: Agentify foundation models with multimodal contextual reasoning, 2026. URL https://arxiv.org/abs/2606.12195

    Pith/arXiv arXiv 2026

  48. [48]

    Momentseeker: A task-oriented benchmark for long-video moment retrieval.Advances in Neural Information Processing Systems, 38, 2026

    Huaying Yuan, Jian Ni, Zheng Liu, Yueze Wang, Junjie Zhou, Zhengyang Liang, Bo Zhao, Zhao Cao, Ji-Rong Wen, and Zhicheng Dou. Momentseeker: A task-oriented benchmark for long-video moment retrieval.Advances in Neural Information Processing Systems, 38, 2026

    2026

  49. [49]

    Tempo-r0: A video-mllm for temporal video grounding through efficient temporal sensing reinforcement learning, 2025

    Feng Yue, Zhaoxing Zhang, Junming Jiao, Zhengyu Liang, Shiwen Cao, Feifei Zhang, and Rong Shen. Tempo-r0: A video-mllm for temporal video grounding through efficient temporal sensing reinforcement learning, 2025. URL https://arxiv.org/abs/2507 .04702

    2025

  50. [50]

    Timesuite: Improving mllms for long video understanding via grounded tuning

    Xiangyu Zeng, Kunchang Li, Chenting Wang, Xinhao Li, Tianxiang Jiang, Ziang Yan, Songze Li, Yansong Shi, Zhengrong Yue, Yi Wang, et al. Timesuite: Improving mllms for long video understanding via grounded tuning. InInternational Conference on Learning Representations, volume 2025, pp. 38057–38081, 2025

    2025

  51. [51]

    Video-o3: Native interleaved clue seeking for long video multi-hop reasoning.arXiv preprint arXiv:2601.23224, 2026

    Xiangyu Zeng, Zhiqiu Zhang, Yuhan Zhu, Xinhao Li, Zikang Wang, Changlian Ma, Qingyu Zhang, Zizheng Huang, Kun Ouyang, Tianxiang Jiang, et al. Video-o3: Native interleaved clue seeking for long video multi-hop reasoning.arXiv preprint arXiv:2601.23224, 2026

    Pith/arXiv arXiv 2026

  52. [52]

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

    Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. InProceedings of the 2023 conference on empirical methods in natural language processing: system demonstrations, pp. 543–553, 2023

    2023

  53. [53]

    Timelens: Rethinking video temporal grounding with multimodal llms

    Jun Zhang, Teng Wang, Yuying Ge, Yixiao Ge, Xinhao Li, and Limin Wang. Timelens: Rethinking video temporal grounding with multimodal llms. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 10419–10429 , 2026

    2026

  54. [54]

    Luowei Zhou, Chenliang Xu, and Jason J. Corso. Towards automatic learning of procedures from web instructional videos, 2017 . URL https://arxiv.org/abs/1703.09788

    Pith/arXiv arXiv 2017

  55. [55]

    Dual detrs for multi- label temporal action detection, 2024

    Yuhan Zhu, Guozhen Zhang, Jing Tan, Gangshan Wu, and Limin Wang. Dual detrs for multi- label temporal action detection, 2024. URL https://arxiv.org/abs/2404.00653

    Pith/arXiv arXiv 2024

  56. [56]

    Freeret: Mllms as training-free retrievers, 2026

    Yuhan Zhu, Xiangyu Zeng, Chenting Wang, Xinhao Li, Chunxu Liu, Yicheng Xu, Ziang Yan, Yi Wang, and Limin Wang. Freeret: Mllms as training-free retrievers, 2026. URL https://arxiv. org/abs/2509 .24621

    2026

  57. [57]

    When doesqhappen?

    Dimitri Zhukov, Jean-Baptiste Alayrac, Ramazan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos, 2019 . URL https://arxiv.org/abs/1903.08225. 16 TimeLens2 Supplementary Material ASupplementary Material Contents A.1 Instruction and Response-Format Diversity . . . . . . . . . . . . ....

    Pith/arXiv arXiv 2019

  58. [2024]

    URL https://arxiv.org/abs/2402.11435

    Pith/arXiv arXiv

  59. [2025]

    URL https://arxiv.org/abs/2511.16150

    arXiv

  60. [2026]

    URL https://arxiv.org/abs/2601.04720

    Pith/arXiv arXiv