Pith. sign in

REVIEW 4 major objections 5 minor 63 references

Uncertainty-quantified Rollout Policy Adaptation for Unlabelled Cross-domain Temporal Grounding

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A test-time adaptation method, URPA, uses the standard deviation across GRPO rollouts as an uncertainty signal to adapt video temporal grounding models to new domains using only 100–200 unlabelled videos.

desk verdict A practical data-efficient adaptation recipe that likely works, wrapped in a theoretical claim that does not. read the letter →

arxiv 2508.06317 v1 pith:AQPOPJ55 submitted 2025-08-08 cs.CV

Jian Hu , Zixu Cheng , Shaogang Gong , Isabel Guan , Jianye Hao , Jun Wang , Kun Shao This is my paper
classification cs.CV
keywords temporalgroundingcross-domainadaptationtest-timeuncertaintyquantificationGRPOpseudo-labelingvision-languagemodelreinforcementlearning
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

The paper proposes URPA, a test-time adaptation method for video temporal grounding that needs only 100–200 unlabelled target-domain videos. It argues that under a GRPO-trained policy, the standard deviation across multiple sampled rollout predictions estimates the Bayesian predictive standard deviation—i.e., epistemic uncertainty—and can therefore be used to weight pseudo-label rewards during adaptation. The central theoretical claim is Theorem 4.1: as the number of rollouts grows and the policy's KL divergence to the source Bayesian predictive distribution shrinks, the rollout standard deviation converges in probability to the true predictive standard deviation. The empirical claim is that this uncertainty-weighted adaptation consistently improves cross-domain grounding over the source-only model, and in several settings matches or beats full-dataset unsupervised adaptation baselines. If true, this gives a label-free, low-data path to adapting large vision-language models to new video domains.

What carries the argument

The central object is the GRPO rollout and its standard deviation. GRPO generates G candidate responses by sampling from the policy πθ; URPA averages the predicted start/end timestamps to form a pseudo-label and computes the rollout standard deviation u. This u is treated as a proxy for epistemic uncertainty, analogous to MC Dropout's variance across subnetworks, and converted into a reward weight via exp(−γu). This mechanism carries the argument because pseudo-label quality is never observed, so the entire adaptation rests on whether rollout variance actually tracks prediction reliability.

What would settle it

Take a handful of labelled target-domain videos, keep them out of adaptation, run URPA, and measure the correlation between rollout standard deviation u and the error of the pseudo-label against the true boundary. If exp(−γu) does not rank pseudo-labels by actual reliability—or if u does not shrink when KL(πθ∥p⋆) is forced small—the central claim that rollout variance quantifies epistemic uncertainty is falsified.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the variance across GRPO rollouts—multiple stochastic temporal-boundary predictions sampled from a policy—acts as a usable epistemic-uncertainty signal for self-training under domain shift. The authors formulate temporal grounding as a reinforcement-learning problem using GRPO, generate G rollouts per target video, average them into a pseudo-label, and compute the standard deviation u of the predicted start and end times. They map u to a confidence weight c = exp(−γu) and multiply the temporal accuracy reward by c during a lightweight gradient update. Theorem 4.1 claims that this rollout standard deviation is a consistent estimator of the Bayesian predic

Load-bearing premise

The theorem's guarantee requires that during unlabelled target adaptation the policy's KL divergence to the source Bayesian predictive distribution tends to zero, yet no ground-truth reward exists in the target domain to drive that convergence.

Editorial extensions

If this is right

  • Using only 100 or 200 unlabelled target videos, URPA improves over source-only GRPO in all six cross-domain settings reported.
  • In several settings (e.g., TACoS→ActivityNet, TACoS→Charades, ActivityNet→Charades), URPA matches or exceeds full-dataset unsupervised domain adaptation baselines.
  • The ablation shows that both the relaxed tIoU and the uncertainty weighting contribute independent gains, with the full URPA variant performing best.
  • Rollout count G=8 is the practical optimum; more rollouts do not help and cost more compute, so the method is efficient as well as data-light.
  • If Theorem 4.1 holds, rollout standard deviation is a principled confidence signal for unlabelled adaptation rather than an ad hoc heuristic.

Reading between the lines

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

  • The convergence claim relies on KL(πθ∥p⋆)→0 during adaptation, but no labelled reward exists in the target domain to drive that convergence; if the policy drifts away, the bound does not apply and the method's success may owe more to variance as a self-consistency measure than to Bayesian uncertainty.
  • A direct testable extension: on held-out labelled target videos, check whether the confidence score exp(−γu) ranks pseudo-labels by their actual boundary error; weak calibration would not necessarily break the method but would show the stated reason is incomplete.
  • The same uncertainty-weighting recipe could be applied to any GRPO-trained structured-output model, but the data-efficiency result may not transfer because temporal grounding has high query redundancy that averages out noise over rollouts.
  • The full-dataset UDA baselines are older task-specific architectures; a stronger future comparison would pit URPA against VLM-based adaptation methods such as entropy minimisation on the same backbone.
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

4 major / 5 minor

Summary. The paper proposes URPA, a data-efficient method for unlabelled cross-domain video temporal grounding. A Qwen2.5-7B policy is first trained on a labelled source domain with supervised GRPO, using a format reward and a relaxed tIoU accuracy reward. At adaptation time, only K=100 or 200 unlabelled target videos are used: the policy samples G rollouts per video, averages the predicted start/end timestamps to form pseudo labels, computes the standard deviation across rollouts as an uncertainty estimate, maps it to a confidence score c=exp(-gamma*u), and weights the tIoU reward by c during a lightweight GRPO update. The paper claims Theorem 4.1: as G tends to infinity and KL(pi_theta||p*) tends to 0, the rollout standard deviation converges in probability to the Bayesian predictive standard deviation, thereby quantifying epistemic uncertainty. Empirically, the paper reports gains over source-only GRPO across six cross-domain settings and provides ablations on Charades->ActivityNet.

Significance. If the claims held, the paper would be a useful practical contribution: a simple, resource-efficient adaptation recipe for temporal grounding with very few unlabelled videos, evaluated on standard benchmarks. The main strengths are the breadth of domain-shift configurations and the ablation showing that uncertainty weighting and relaxed pseudo-label tIoU both help. However, the theoretical justification in Section 4 is not currently established: the proof relies on an unsupported assumption that GRPO adaptation drives the policy toward the source Bayesian predictive distribution, while the target adaptation objective in Section 3.4 has no mechanism to ensure this. The empirical evaluation also lacks repeated-seed statistics, which matters for randomly selected 100/200-sample subsets. The practical recipe may still be useful, but the uncertainty-quantification interpretation requires either a corrected conditional theorem or direct empirical validation.

major comments (4)
  1. [Section 4, Theorem 4.1 Step 3] The theorem's conclusion depends on epsilon=KL(pi_theta||p*) -> 0, but Step 3 simply asserts that 'GRPO optimisation decreases epsilon over training'. For the target adaptation actually deployed, the objective in Eq. (10) is computed from pseudo labels generated by averaging the policy's own rollouts (Eq. 7) and confidence weights derived from the same rollouts (Eq. 9). The KL regularization in Eq. (2) is with respect to pi_ref, not p*, and there is no term in the adaptation objective that drives pi_theta toward the source Bayesian predictive distribution p*. Self-training on a shifted, unlabelled domain could plausibly increase epsilon. Without a proof or empirical evidence that epsilon->0, the chain in Eqs. (13)-(16) does not connect rollout standard deviation to epistemic uncertainty in the deployed setting. The theorem should be restated as conditional, or the proof must supply a mec
  2. [Section 4, proof of Theorem 4.1] The proof refers to 'supplementary material Sec.??' for full details, but no supplementary section is present in the submitted manuscript. The central variance-gap bound in Eq. (13) is asserted after Pinsker's inequality and the bounded-moment assumption without a derivation. Similarly, Section 5.2 refers to an 'Appendix??' for data-efficient supervised baselines. These missing references prevent independent verification of the paper's central theoretical claim and of part of the empirical comparison. The missing material must be included, or the references removed and the key steps shown in the main text.
  3. [Section 3.4, Eqs. (7)-(10)] The confidence weighting is circular in an important sense: the pseudo labels and confidence scores are both produced by the very policy being adapted. Low rollout variance may reflect systematic overconfidence under domain shift rather than low epistemic uncertainty, and the theorem's p* is defined with respect to source data. The only evidence that uncertainty weighting helps is the single ablation row in Table 2, which shows a gain on one transfer task but does not establish that rollout standard deviation is calibrated to pseudo-label error on the target domain. A direct validation--for example, the correlation between u and prediction error on held-out target samples, or a comparison against an MC-Dropout uncertainty estimate--would substantiate the uncertainty interpretation.
  4. [Table 1] The central empirical claim is that URPA 'consistently outperforms' source-only GRPO across six domain-shift settings with K=100 or 200 unlabelled target videos. Table 1 reports single numbers without error bars or multiple seeds. Since the target videos are randomly selected subsets of size 100/200, the reported gains (often 1-3 R@ points) may fall within sampling noise. The authors should report at least three independent runs with means and standard deviations, or provide statistical significance tests, particularly for the 100-shot results.
minor comments (5)
  1. [Abstract and Section 1] The abstract contains an incomplete sentence: 'Specifically, we introduce.' followed by a new sentence. There is also a grammatical error in the introduction: 'requiring extensively labelled training videos' should be 'requiring extensive labelled training videos'.
  2. [Section 3.4, Eq. (10)] The symbol R_s is reused for both the source supervised reward (Eq. 6) and the target adaptation reward (Eq. 10). Rename the latter R_t to avoid confusion.
  3. [Table 2] Row 2 says 'Our GRPA Target Adaptation' but the method is called URPA. Also, applying the target-adaptation procedure to a base Qwen2.5-7B without source GRPO training is not clearly a valid instance of URPA, since URPA assumes a rollout policy already trained for grounding. Please clarify what this row actually evaluates.
  4. [Section 5.2, Fig. 2] Figure 2 contains garbled mathematical characters and overlapped labels, making the diagram difficult to read. Please regenerate the figure with clean typesetting.
  5. [Section 3.4, Eq. (8)] The uncertainty u is defined as the sum of standard deviations of start and end timestamps. The choice to sum rather than use a joint covariance or vector norm is not discussed. Please justify this modeling choice or show that the results are insensitive to it.

Circularity Check

1 steps flagged · score 4.0 of 10

Theorem 4.1's Step 3 assumes KL(pi_theta || p*) -> 0 during a target adaptation whose only reward signal is built from the current policy's own rollouts, so the epistemic-uncertainty interpretation of rollout variance is conditioned on an unproven premise; held-out Table 1 evaluations keep the empirical claims independent.

  1. other [Section 4, Theorem 4.1, Proof Step 3 (Eqs. 13-16)]
    "Step 3: Convergence to Bayesian predictive standard deviation. GRPO optimisation decreases ε over training, so √ε→0. Combining (11) and the above yields ... which tends to zero in probability as G→∞ and ε→0. Hence bσG(x) −→p σ⋆(x) ... showing that the rollout standard deviation is a consistent estimator of epistemic uncertainty."

    The theorem's conclusion — rollout std bσG(x) converges to the Bayesian predictive std σ⋆(x), 'showing that the rollout standard deviation is a consistent estimator of epistemic uncertainty' — holds only under Step 3's assertion that GRPO drives ε=KL(πθ∥p⋆)→0. In the deployed URPA target adaptation, the reward (Eq. 10) is computed from pseudo-labels that are the mean of the same policy's rollouts (Eq. 7) and a confidence score that is the std of those same rollouts (Eqs. 8-9); no term involves p⋆, and the paper gives no argument that self-training on shifted unlabelled videos moves πθ toward the source-domain Bayesian predictive distribution. When ε=0 the bound (13) is vacuous and σπ=σ⋆ by definition, so the claimed 'prediction' that rollout variance equals epistemic uncertainty reduces to

full rationale

The empirical contribution is externally validated: URPA adapts on 100-200 unlabelled target videos and is evaluated on held-out target test sets (Table 1), so the reported gains are not produced by fitting the test data and do not reduce to the training objective by construction. Self-citations ([3, 18, 19, 20]) appear only in related-work or backbone contexts and are not load-bearing; the MC Dropout analogy relies on the external, standard results of Gal & Ghahramani, not on a self-citation chain. No uniqueness theorem is imported, and no known result is simply renamed. The one substantive issue is the theoretical claim (contribution iii). The proof of Theorem 4.1 is a valid conditional derivation (LLN in Step 1, Pinsker in Step 2), but its load-bearing Step 3 asserts without derivation that 'GRPO optimisation decreases ε over training.' In the actual target-adaptation phase, the objective (Eq. 10) is entirely a function of the current policy's rollouts: pseudo-label = rollout mean (Eq. 7), confidence = exp(-γ * rollout std) (Eqs. 8-9). Nothing in that self-referential objective references the source-domain Bayesian predictive p⋆, and the paper supplies no mechanism or evidence that this objective drives πθ toward p⋆. Consequently the theorem's conclusion (16) and the resulting claim that rollout variance quantifies epistemic uncertainty are conditioned on an unproven premise that the deployed setting does not establish; the proof's own details are deferred to a missing 'Sec.??' and are not verifiable. This is partial circularity — the theory's headline result rests on assuming the very convergence it needs — but because the empirical results stand independently of the theorem, the paper is not wholly circular. Score 4.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The method relies on three fitted hyperparameters (alpha, gamma, G) and on unproven modelling assumptions connecting rollout variance to Bayesian epistemic uncertainty. No new physical entities are introduced; the confidence score is a deterministic function of rollouts, not a postulated entity.

free parameters (4)
  • relaxation ratio alpha = 0.1
    Eq.(4) extends ground-truth and pseudo-label boundaries by alpha times event duration; fixed by hand with no sensitivity analysis.
  • uncertainty decay gamma = 10
    Eq.(9) converts rollout std into confidence; Table 3(a) shows gamma=10 best on Charades->ActivityNet, so it is tuned on a target task.
  • rollout count G = 8
    Number of rollouts used to form pseudo labels and variance; Table 3(b) shows G=8 peaks on one transfer task, larger G lowers R@0.5 and mIoU.
  • reward weights for format and tIoU = 0.5 / 0.5
    Eq.(6) and Eq.(10) fix equal weights for format and groundedness rewards; not justified by calibration.
assumptions (5)
  • domain assumption Rollouts from pi_theta are i.i.d.
    Assumption (i) of Theorem 4.1; stochastic sampling is treated as independent draws from the same policy for a fixed input.
  • domain assumption There is a Bayesian predictive distribution p* for the same model class trained on source data, with finite second moments and finite KL.
    Assumptions (ii)-(iii) of Theorem 4.1; the existence and relevance of p* to the adapted target policy is not established.
  • ad hoc to paper GRPO training drives KL(pi_theta || p*) to zero.
    Proof Step 3 of Theorem 4.1 asserts this convergence, but neither source RL nor target pseudo-label adaptation is shown to minimize KL to a Bayesian posterior.
  • ad hoc to paper Low rollout variance implies reliable pseudo labels, so confidence weighting improves adaptation.
    Core design choice in Section 3.4 Eq.(8)-(10); no calibration experiment links variance to pseudo-label error, and variance can shrink from policy collapse.
  • standard math Pinsker's inequality and the weak law of large numbers apply.
    Used in Theorem 4.1 proof; standard but require integrability and absolute continuity not stated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Uncertainty-quantified Rollout Policy Adaptation for Unlabelled Cross-domain Temporal Grounding." pith.science (2026). https://pith.science/paper/AQPOPJ55

@misc{pith2026250806317,
  author       = {Pith},
  title        = {Pith review of: Uncertainty-quantified Rollout Policy Adaptation for Unlabelled Cross-domain Temporal Grounding},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AQPOPJ55}},
  note         = {Machine review of arXiv:2508.06317}
}
read the original abstract

Video Temporal Grounding (TG) aims to temporally locate video segments matching a natural language description (a query) in a long video. While Vision-Language Models (VLMs) are effective at holistic semantic matching, they often struggle with fine-grained temporal localisation. Recently, Group Relative Policy Optimisation (GRPO) reformulates the inference process as a reinforcement learning task, enabling fine-grained grounding and achieving strong in-domain performance. However, GRPO relies on labelled data, making it unsuitable in unlabelled domains. Moreover, because videos are large and expensive to store and process, performing full-scale adaptation introduces prohibitive latency and computational overhead, making it impractical for real-time deployment. To overcome both problems, we introduce a Data-Efficient Unlabelled Cross-domain Temporal Grounding method, from which a model is first trained on a labelled source domain, then adapted to a target domain using only a small number of unlabelled videos from the target domain. This approach eliminates the need for target annotation and keeps both computational and storage overhead low enough to run in real time. Specifically, we introduce. Uncertainty-quantified Rollout Policy Adaptation (URPA) for cross-domain knowledge transfer in learning video temporal grounding without target labels. URPA generates multiple candidate predictions using GRPO rollouts, averages them to form a pseudo label, and estimates confidence from the variance across these rollouts. This confidence then weights the training rewards, guiding the model to focus on reliable supervision. Experiments on three datasets across six cross-domain settings show that URPA generalises well using only a few unlabelled target videos. Codes will be released once published.

Figures

Figures reproduced from arXiv: 2508.06317 by the authors.

Figure 1
Figure 1. (a) A comparison between full-data and data-efficient adaptation in Cross-domain Temporal [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Uncertainty-quantified Rollout Policy Adaptation (URPA): During source model training, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Qualitative Analysis on Charades → ActivityNet. approach mitigates biases introduced by manual annotations during pre-training. The last three rows analyse different modules in our target domain adaptation strategy, starting from a source-trained model. In the fifth row, pseudo labels are used directly for adaptation without modification in Eq.(4) and uncertainty quantification in Eq.(10), resulting in the lowest pe… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

63 extracted references · 36 canonical work pages

  1. [1]

    Localizing moments in video with natural language

    Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. InProceedings of the IEEE 9 international conference on computer vision, pages 5803–5812, 2017

    work page 2017

  2. [2]

    Activitynet: A large-scale video benchmark for human activity understanding

    Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. InProceedings of the ieee conference on computer vision and pattern recognition, pages 961–970, 2015

    2015

  3. [3]

    Shine: Saliency- aware hierarchical negative ranking for compositional temporal grounding

    Zixu Cheng, Yujiang Pu, Shaogang Gong, Parisa Kordjamshidi, and Yu Kong. Shine: Saliency- aware hierarchical negative ranking for compositional temporal grounding. InEuropean Conference on Computer Vision, pages 398–416. Springer, 2024

    work page 2024

  4. [4]

    Scaling instruction-finetuned language models.Journal of Machine Learning Research, 25(70):1–53, 2024

    Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. Scaling instruction-finetuned language models.Journal of Machine Learning Research, 25(70):1–53, 2024

    2024

  5. [5]

    Aleatory or epistemic? does it matter?Structural safety, 31(2):105–112, 2009

    Armen Der Kiureghian and Ove Ditlevsen. Aleatory or epistemic? does it matter?Structural safety, 31(2):105–112, 2009

    2009

  6. [6]

    Partially relevant video retrieval

    Jianfeng Dong, Xianke Chen, Minsong Zhang, Xun Yang, Shujie Chen, Xirong Li, and Xun Wang. Partially relevant video retrieval. InProceedings of the 30th ACM International Conference on Multimedia, pages 246–257, 2022

    work page 2022

  7. [7]

    Dual encoding for video retrieval by text.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):4065–4080, 2021

    Jianfeng Dong, Xirong Li, Chaoxi Xu, Xun Yang, Gang Yang, Xun Wang, and Meng Wang. Dual encoding for video retrieval by text.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):4065–4080, 2021

    work page 2021

  8. [8]

    Not all inputs are valid: Towards open-set video moment retrieval using language

    Xiang Fang, Wanlong Fang, Daizong Liu, Xiaoye Qu, Jianfeng Dong, Pan Zhou, Renfu Li, Zichuan Xu, Lixing Chen, Panpan Zheng, et al. Not all inputs are valid: Towards open-set video moment retrieval using language. InProceedings of the 32nd ACM International Conference on Multimedia, pages 28–37, 2024

    work page 2024

Show all 63 references
  1. [9]

    Detecting hallucinations in large language models using semantic entropy.Nature, 630(8017):625–630, 2024

    Sebastian Farquhar, Jannik Kossen, Lorenz Kuhn, and Yarin Gal. Detecting hallucinations in large language models using semantic entropy.Nature, 630(8017):625–630, 2024

    2024

  2. [10]

    Gal and Z

    Y . Gal and Z. Ghahramani. Dropout as a bayesian approximation: Representing model uncer- tainty in deep learning. InICML, 2016

    2016

  3. [11]

    Dropout as a bayesian approximation: Representing model uncertainty in deep learning

    Yarin Gal and Zoubin Ghahramani. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. Ininternational conference on machine learning, pages 1050–1059. PMLR, 2016

    2016

  4. [12]

    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, pages 5267–5275, 2017

    2017

  5. [13]

    On calibration of modern neural networks

    Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. InInternational conference on machine learning, pages 1321–1330. PMLR, 2017

    2017

  6. [14]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    2025 arXiv

  7. [15]

    Vtg-llm: Integrating timestamp knowledge into video llms for enhanced video temporal grounding

    Yongxin Guo, Jingyu Liu, Mingda Li, Dingxin Cheng, Xiaoying Tang, Dianbo Sui, Qingbin Liu, Xi Chen, and Kevin Zhao. Vtg-llm: Integrating timestamp knowledge into video llms for enhanced video temporal grounding. InProceedings of the AAAI Conference on Artificial Intelligence, ...

    2025

  8. [16]

    Trace: Temporal grounding video llm via causal event modeling.arXiv preprint arXiv:2410.05643, 2024

    Yongxin Guo, Jingyu Liu, Mingda Li, Qingbin Liu, Xi Chen, and Xiaoying Tang. Trace: Temporal grounding video llm via causal event modeling.arXiv preprint arXiv:2410.05643, 2024

    2024 arXiv

  9. [17]

    Cone: An efficient coarse-to-fine alignment framework for long video temporal grounding

    Zhijian Hou, Wanjun Zhong, Lei Ji, Difei Gao, Kun Yan, Wk Chan, Chong-Wah Ngo, Mike Zheng Shou, and Nan Duan. Cone: An efficient coarse-to-fine alignment framework for long video temporal grounding. InProceedings of the 61st Annual Meeting of the Association for Computational ...

    2023

  10. [18]

    Cos: Chain-of-shot prompting for long video understanding.arXiv preprint arXiv:2502.06428, 2025

    Jian Hu, Zixu Cheng, Chenyang Si, Wei Li, and Shaogang Gong. Cos: Chain-of-shot prompting for long video understanding.arXiv preprint arXiv:2502.06428, 2025

    2025 arXiv

  11. [19]

    Discriminative partial domain adversarial network

    Jian Hu, Hongya Tuo, Chao Wang, Lingfeng Qiao, Haowen Zhong, Junchi Yan, Zhongliang Jing, and Henry Leung. Discriminative partial domain adversarial network. InComputer Vision– ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXVII 16, pa...

    2020

  12. [20]

    Learning unbiased transferability for domain adaptation by uncertainty modeling

    Jian Hu, Haowen Zhong, Fei Yang, Shaogang Gong, Guile Wu, and Junchi Yan. Learning unbiased transferability for domain adaptation by uncertainty modeling. InEuropean Conference on Computer Vision, pages 223–241. Springer, 2022

    2022

  13. [21]

    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, pages 14271–14280, 2024

    2024

  14. [22]

    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, pages 13700–1...

    2024

  15. [23]

    What uncertainties do we need in bayesian deep learning for computer vision?arXiv preprint arXiv:1703.04977, 2017

    Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision?arXiv preprint arXiv:1703.04977, 2017

    2017 arXiv

  16. [24]

    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, pages 706–715, 2017

    2017

  17. [25]

    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. [26]

    Envolving temporal reasoning capability into lmms via temporal consistent reward

    Hongyu Li, Songhao Han, Yue Liao, Jialin Gao, and Si Liu. Envolving temporal reasoning capability into lmms via temporal consistent reward. https://github.com/appletea233/ Temporal-R1, 2025

    2025

  19. [27]

    Compositional temporal grounding with structured variational cross-graph correspondence learning

    Juncheng Li, Junlin Xie, Long Qian, Linchao Zhu, Siliang Tang, Fei Wu, Yi Yang, Yueting Zhuang, and Xin Eric Wang. Compositional temporal grounding with structured variational cross-graph correspondence learning. InProceedings of the IEEE/CVF Conference on Computer Vision and ...

    2022

  20. [28]

    Groundinggpt: Language enhanced multi-modal grounding model

    Zhaowei Li, Qi Xu, Dong Zhang, Hang Song, YiQing Cai, Qi Qi, Ran Zhou, Junting Pan, Zefeng Li, Vu Tu, et al. Groundinggpt: Language enhanced multi-modal grounding model. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long P...

    2024

  21. [29]

    Weakly-supervised video moment retrieval via semantic completion network

    Zhijie Lin, Zhou Zhao, Zhu Zhang, Qi Wang, and Huasheng Liu. Weakly-supervised video moment retrieval via semantic completion network. InProceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 11539–11546, 2020

    2020

  22. [30]

    Unsupervised domain adaptative temporal sentence localization with mutual in- formation maximization

    Daizong Liu, Xiang Fang, Xiaoye Qu, Jianfeng Dong, He Yan, Yang Yang, Pan Zhou, and Yu Cheng. Unsupervised domain adaptative temporal sentence localization with mutual in- formation maximization. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages...

    2024

  23. [31]

    Et bench: Towards open-ended event-level video-language understanding.Advances in Neural Information Processing Systems, 37:32076–32110, 2024

    Ye Liu, Zongyang Ma, Zhongang Qi, Yang Wu, Ying Shan, and Chang W Chen. Et bench: Towards open-ended event-level video-language understanding.Advances in Neural Information Processing Systems, 37:32076–32110, 2024

    2024

  24. [32]

    The norm must go on: Dynamic unsupervised domain adaptation by normalization

    M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14765–14775, 2022. 11

    2022

  25. [33]

    Query- dependent video representation for moment retrieval and highlight detection

    WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query- dependent video representation for moment retrieval and highlight detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 23023–23033, 2023

    2023

  26. [34]

    Local-global video-text interactions for temporal grounding

    Jonghwan Mun, Minsu Cho, and Bohyung Han. Local-global video-text interactions for temporal grounding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10810–10819, 2020

    2020

  27. [35]

    Efficient test-time model adaptation without forgetting

    Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. InInternational confer- ence on machine learning, pages 16888–16905. PMLR, 2022

    2022

  28. [36]

    Pawlowski, A

    N. Pawlowski, A. Brock, Matthew C. H Lee, M. Rajchl, and B. Glocker. Implicit weight uncertainty in neural networks.arXiv preprint arXiv:1711.01297, 2017

    2017 arXiv

  29. [37]

    Grounding is all you need? dual temporal grounding for video dialog.arXiv preprint arXiv:2410.05767, 2024

    You Qin, Wei Ji, Xinze Lan, Hao Fei, Xun Yang, Dan Guo, Roger Zimmermann, and Lizi Liao. Grounding is all you need? dual temporal grounding for video dialog.arXiv preprint arXiv:2410.05767, 2024

    2024 arXiv

  30. [38]

    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. [39]

    Timechat: A time-sensitive multimodal large language model for long video understanding

    Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14313–14323, 2024

    2024

  32. [40]

    Hollywood in homes: Crowdsourcing data collection for activity understanding

    Gunnar A Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. InComputer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Pro...

    2016

  33. [41]

    Video understanding with large language models: A survey.IEEE Transactions on Circuits and Systems for Video Technology, 2025

    Yunlong Tang, Jing Bi, Siting Xu, Luchuan Song, Susan Liang, Teng Wang, Daoan Zhang, Jie An, Jingyang Lin, Rongyi Zhu, et al. Video understanding with large language models: A survey.IEEE Transactions on Circuits and Systems for Video Technology, 2025

    2025

  34. [42]

    Tent: Fully test-time adaptation by entropy minimization.arXiv preprint arXiv:2006.10726, 2020

    Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization.arXiv preprint arXiv:2006.10726, 2020

    2006 arXiv

  35. [43]

    Routing evidence for unseen actions in video moment retrieval

    Guolong Wang, Xun Wu, Zheng Qin, and Liangliang Shi. Routing evidence for unseen actions in video moment retrieval. InProceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining, pages 3024–3035, 2024

    2024

  36. [44]

    Temporally grounding language queries in videos by contextual boundary-aware prediction

    Jingwen Wang, Lin Ma, and Wenhao Jiang. Temporally grounding language queries in videos by contextual boundary-aware prediction. InProceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12168–12175, 2020

    2020

  37. [45]

    Dynamically instance-guided adaptation: A backward-free approach for test-time domain adaptive semantic segmentation

    Wei Wang, Zhun Zhong, Weijie Wang, Xi Chen, Charles Ling, Boyu Wang, and Nicu Sebe. Dynamically instance-guided adaptation: A backward-free approach for test-time domain adaptive semantic segmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Rec...

    2023

  38. [46]

    Transferable calibration with lower bias and variance in domain adaptation.arXiv preprint arXiv:2007.08259, 2020

    Ximei Wang, Mingsheng Long, Jianmin Wang, and Michael I Jordan. Transferable calibration with lower bias and variance in domain adaptation.arXiv preprint arXiv:2007.08259, 2020

    2007 arXiv

  39. [47]

    Timezero: Temporal video grounding with reasoning-guided lvlm

    Ye Wang, Boshen Xu, Zihao Yue, Zihan Xiao, Ziheng Wang, Liang Zhang, Dingyi Yang, Wenxuan Wang, and Qin Jin. Timezero: Temporal video grounding with reasoning-guided lvlm. arXiv preprint arXiv:2503.13377, 2025

    2025 arXiv

  40. [48]

    Negative sample matters: A renaissance of metric learning for temporal grounding

    Zhenzhi Wang, Limin Wang, Tao Wu, Tianhao Li, and Gangshan Wu. Negative sample matters: A renaissance of metric learning for temporal grounding. InProceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2613–2623, 2022. 12

    2022

  41. [49]

    Can llms express their uncertainty? an empirical evaluation of confidence elicitation in llms.arXiv preprint arXiv:2306.13063, 2023

    Miao Xiong, Zhiyuan Hu, Xinyang Lu, Yifei Li, Jie Fu, Junxian He, and Bryan Hooi. Can llms express their uncertainty? an empirical evaluation of confidence elicitation in llms.arXiv preprint arXiv:2306.13063, 2023

    2023 arXiv

  42. [50]

    Large language models can learn temporal reasoning.arXiv preprint arXiv:2401.06853, 2024

    Siheng Xiong, Ali Payani, Ramana Kompella, and Faramarz Fekri. Large language models can learn temporal reasoning.arXiv preprint arXiv:2401.06853, 2024

    2024 arXiv

  43. [51]

    Multilevel language and vision integration for text-to-clip retrieval

    Huijuan Xu, Kun He, Bryan A Plummer, Leonid Sigal, Stan Sclaroff, and Kate Saenko. Multilevel language and vision integration for text-to-clip retrieval. InProceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9062–9069, 2019

    2019

  44. [52]

    Unloc: A unified framework for video localization tasks

    Shen Yan, Xuehan Xiong, Arsha Nagrani, Anurag Arnab, Zhonghao Wang, Weina Ge, David Ross, and Cordelia Schmid. Unloc: A unified framework for video localization tasks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13623– 13633, 2023

    2023

  45. [53]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report.arXiv preprint arXiv:2412.15115, 2024

    2024 arXiv

  46. [54]

    A closer look at temporal sentence grounding in videos: Dataset and metric

    Yitian Yuan, Xiaohan Lan, Xin Wang, Long Chen, Zhi Wang, and Wenwu Zhu. A closer look at temporal sentence grounding in videos: Dataset and metric. InProceedings of the 2nd international workshop on human-centric multimedia analysis, pages 13–21, 2021

    2021

  47. [55]

    Semantic conditioned dynamic modulation for temporal sentence grounding in videos.Advances in Neural Information Processing Systems, 32, 2019

    Yitian Yuan, Lin Ma, Jingwen Wang, Wei Liu, and Wenwu Zhu. Semantic conditioned dynamic modulation for temporal sentence grounding in videos.Advances in Neural Information Processing Systems, 32, 2019

    2019

  48. [56]

    To find where you talk: Temporal sentence localization in video with attention based location regression

    Yitian Yuan, Tao Mei, and Wenwu Zhu. To find where you talk: Temporal sentence localization in video with attention based location regression. InProceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 9159–9166, 2019

    2019

  49. [57]

    Dense regression network for video grounding

    Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287–10296, 2020

    2020

  50. [58]

    Timesuite: Improving mllms for long video understanding via grounded tuning.arXiv preprint arXiv:2410.19702, 2024

    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.arXiv preprint arXiv:2410.19702, 2024

    2024 arXiv

  51. [59]

    Span-based localizing network for natural language video localization

    Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Span-based localizing network for natural language video localization. InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6543–6554, 2020

    2020

  52. [60]

    No-frills temporal video grounding: Multi-scale neighboring attention and zoom-in boundary detection.arXiv preprint arXiv:2307.10567, 2023

    Qi Zhang, Sipeng Zheng, and Qin Jin. No-frills temporal video grounding: Multi-scale neighboring attention and zoom-in boundary detection.arXiv preprint arXiv:2307.10567, 2023

    2023 arXiv

  53. [61]

    Learning 2d temporal adjacent networks for moment localization with natural language

    Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. InProceedings of the AAAI conference on artificial intelligence, volume 34, pages 12870–12877, 2020

    2020

  54. [62]

    Cross-modal interaction networks for query-based moment retrieval in videos

    Zhu Zhang, Zhijie Lin, Zhou Zhao, and Zhenxin Xiao. Cross-modal interaction networks for query-based moment retrieval in videos. InProceedings of the 42nd international ACM SIGIR conference on research and development in information retrieval, pages 655–664, 2019

    2019

  55. [63]

    Verify-and-edit: A knowledge-enhanced chain-of-thought framework.arXiv preprint arXiv:2305.03268, 2023

    Ruochen Zhao, Xingxuan Li, Shafiq Joty, Chengwei Qin, and Lidong Bing. Verify-and-edit: A knowledge-enhanced chain-of-thought framework.arXiv preprint arXiv:2305.03268, 2023. 13

    2023 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.