Pith. sign in

REVIEW 5 major objections 6 minor 2 cited by

Tempo-R0: A Video-MLLM for Temporal Video Grounding through Efficient Temporal Sensing Reinforcement Learning

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Tempo-R0 claims that a video-LLM pipeline with refusal-aware reinforcement outperforms state-of-the-art temporal grounding by about 3.5% on QVHighlights.

desk verdict Plausible RLVR recipe for temporal grounding, but the headline SOTA gain is confounded by backbone mismatch and an author-corrected testbench. read the letter →

arxiv 2507.04702 v1 pith:CGRJCBEZ submitted 2025-07-07 cs.CV cs.AI

Feng Yue , Zhaoxing Zhang , Junming Jiao , Zhengyu Liang , Shiwen Cao , Feifei Zhang , Rong Shen This is my paper
classification cs.CVcs.AI
keywords TemporalVideoGroundingMultimodalLargeLanguageModelReinforcementFine-TuningGRPOOptimalTransportFrameSamplingNegativeQueryQVHighlights
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

Temporal video grounding asks a model to locate the segment of a video that matches a natural-language query. The paper argues that video-MLLMs struggle with this because they waste limited context on redundant frames, cannot pin timestamps to events, and have no training signal for saying 'no match'. Tempo-R0 answers with three additions: a frame-salience sampler that gives more visual tokens to content-changing frames, explicit fixed-width timestamps interleaved as a modality, and a GRPO reward variant that also rewards refusing irrelevant video-query pairs. On QVHighlights and a manually corrected version, the authors report roughly 3.5% gains over prior state-of-the-art Video-MLLM solutions, with the best metrics reaching 78.52 R@.5 on the original benchmark.

What carries the argument

The argument rides on three mechanisms plus a benchmark correction. SAA computes a hue-channel optimal-transport difference between consecutive sampled frames and gives more visual tokens to frames with large content changes, effectively redistributing the MLLM's limited visual-token budget to boundary frames. ETA converts each sampled frame's floating-point timestamp into zero-padded text of equal token length and interleaves it with visual tokens, so the model treats timestamps as an explicit third modality. PIR-GRPO modifies the GRPO precision reward so that a model receives reward not only for IoU between predicted and ground-truth intervals on relevant pairs but also for outputting no interval on pure-irrelevant pairs; training mixes roughly 10% irrelevant pairs. The corrected cQvH testbench is produced by manual verification of QVHighlights annotations and is used alongside the original.

What would settle it

Run a controlled ablation: take Qwen2-VL-7B, apply the same SFT and GRPO training data but replace SAA with uniform frame sampling, drop ETA, and drop the refusal term; if the resulting model still scores around 78.52 R@.5 on QVHighlights, the pipeline's components are not the cause of the gain.

Watch

Extended reading notes

Core claim

The central claim is that temporal grounding performance of a video-MLLM can be improved without changing the backbone by attacking three bottlenecks: selective allocation of visual tokens to information-rich frames via an optimal-transport hue-based salience score (SAA), explicit timestamp alignment that tokenizes timestamps in equal-length text (ETA), and a two-stage reinforced fine-tuning (SFT then PIR-GRPO) whose reward function includes a refusal term for irrelevant pairs. With a Qwen2-VL-7B backbone, this yields R@.5/R@.7/mAP of 78.52/65.23/54.50 on QVHighlights and 84.65/72.39/61.47 on the corrected cQvH set, exceeding prior models. The paper also contributes a manually corrected QVHighlights test set, arguing its ground truth is more consistent with human perception.

Load-bearing premise

The headline comparison assumes the reported gain comes from the method rather than from Qwen2-VL-7B being a newer and stronger backbone than the BLIP-2-based baselines.

Editorial extensions

If this is right

  • If correct, the reported QVHighlights numbers (78.52/65.23/54.50) exceed Chrono's (74.77/60.51/51.37) on all three metrics, so Tempo-R0 would become the new reference point for Video-MLLM temporal grounding.
  • On the corrected cQvH set the same model reaches 84.65 R@.5, suggesting a substantial share of prior evaluation errors may be annotation artifacts rather than model failures.
  • The refusal-aware reward extends RLVR-style training to negative or unanswerable queries, making abstention an explicit, optimizable behavior for grounding models.
  • Ablations indicate SAA and ETA each add measurable points, and PIR-GRPO adds roughly 5 R@.5 over plain SFT, with a 10% irrelevance ratio performing best.
  • Zero-shot transfer from QVHighlights to ActivityNet (32.22 R@.5) suggests the learned temporal reasoning generalizes beyond the training distribution.

Reading between the lines

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

  • The same refusal-reward idea could be transplanted to other vision-language retrieval tasks with unanswerable queries, such as referring expression comprehension or video question answering, where abstention is currently under-rewarded.
  • A like-for-like backbone control—Qwen2-VL-7B fine-tuned without SAA, ETA, and PIR-GRPO—would separate method gains from backbone gains; the paper's tables compare against BLIP-2-based models, so this control is missing.
  • If cQvH is released with inter-annotator agreement and explicit correction rules, it could become a standard secondary benchmark; without those, annotation bias cannot be ruled out.
  • The OT-based salience criterion may be sensitive to hue changes that are not semantic, such as lighting shifts or camera shake, so testing SAA under such perturbations would reveal whether its advantage is content-based or low-level.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The manuscript proposes Tempo-R0, a video MLLM based on Qwen2-VL-7B for temporal video grounding (TVG). It introduces three components: Self-adaptive Attention Allocation (SAA), which uses optimal-transport-based frame-difference scores to allocate more visual tokens to high-change frames; Explicit Timestamp Alignment (ETA), which interleaves zero-padded textual timestamps with visual tokens; and Partial Irrelevance Refusing GRPO (PIR-GRPO), a two-stage SFT-plus-RLVR procedure that adds pure irrelevant video-query pairs to teach the model to refuse unmatched queries. The paper also presents a manually corrected version of the QVHighlights test set, cQvH. The headline claim is that Tempo-R0 outperforms prior SOTA solutions by about 3.5% on both QVHighlights and cQvH, with the main results reporting QVHighlights R@.5/R@.7/mAP of 78.52/65.23/54.50 against Chrono's 74.77/60.51/51.37, and cQvH values of 84.65/72.39/61.47.

Significance. If the reported gains are reproducible and attributable to the method rather than to the backbone, the paper would make a useful contribution to TVG: SAA's content-adaptive token allocation is a sensible response to context-length limits, ETA is a clean way to make timestamps machine-readable, and PIR-GRPO is a reasonable extension of RLVR to refusal behavior. The corrected cQvH benchmark, if rigorously constructed and released, could be a useful community asset. The ablations cover each proposed component and the paper is generally clear in its motivation. However, the central claim requires a same-backbone comparison against an existing Qwen2-VL-7B TVG method, and the cQvH benchmark needs a transparent validation protocol; the current evidence does not yet establish the stated SOTA advantage. The paper does not provide code, model weights, or the corrected annotations.

major comments (5)
  1. [Main Results (Table 1)] The central claim of a roughly 3.5% advantage over SOTA is not supported by Table 1 as printed, because all three listed comparators (Chrono, SeViLa, InternVideo2) use the older BLIP-2 backbone, whereas Tempo-R0 uses Qwen2-VL-7B; no Qwen2-VL-7B TVG baseline such as Time-R1 (cited as Wang et al. 2025) appears in the QVHighlights rows. The row labeled 'Tempo-R1' in Table 2 is presumably a typo for Time-R1 and reports only ActivityNet transfer numbers, so it does not resolve the comparison. Please add Time-R1's QVHighlights scores, or state explicitly that Time-R1 has not published QVHighlights numbers, and include an SFT-only Qwen2-VL-7B baseline so that the claimed gain can be attributed to the proposed pipeline rather than to the backbone.
  2. [Testbench with More Reasonable Ground Truth (GT) Annotations] The cQvH testbench is manually corrected by the authors with no inter-annotator agreement measure, no pre-specified correction protocol, and no statement that the corrections were made blind to model predictions. Since cQvH is one of the two headline evaluation sets and is used to support the ~3.5% advantage claim, the paper must release the correction protocol, annotation statistics, and an independent human-agreement estimate, and should demonstrate that the corrections are not biased by the authors' own model outputs. Without this, scores on cQvH cannot be treated as evidence of a method-level improvement over the original benchmark.
  3. [Ablation Experiments (Tables 6-7)] Table 6 does not isolate PIR-GRPO from ordinary GRPO: the 'Baseline with SFT' and '+OOD' rows are SFT conditions, and no '+OOD + GRPO with 0% irrelevant pairs' row is reported. Table 7's QvH(0%) row reports exactly the SFT baseline numbers, which suggests that 0% corresponds to no RL at all, so the improvement in '+OOD +PIR-GRPO' could come from RL itself rather than from refusing irrelevant pairs. Please add a plain-GRPO condition with the same data and reward except for the irrelevant-pair term, and clarify the relationship between Tables 6 and 7. Note also that Table 7 QvH(10%) reports 78.84/64.65/54.22, which does not match Table 6's '+OOD +PIR-GRPO' row of 78.52/65.23/54.50.
  4. [Ablation Experiments (Tables 3-5)] The ablation tables do not state which other components are active in each row, and the numbers are mutually inconsistent with the final model: Table 3 'w/ SAA' is 77.39/63.52/53.84, Table 5 'w/ ETA' is 76.77/63.61/53.33, and the full model in Table 1 is 78.52/65.23/54.50; moreover Table 5 'w/ ETA' duplicates Table 3 'w/o SAA' exactly, while Table 6's SFT baseline is much lower. Without a stated common base configuration, the individual contribution of each component cannot be assessed, and Table 3 shows a mixed result (R@.7 decreases from 63.61 to 63.52 with SAA). Please specify the base model for each ablation, report variance across runs, and reconcile the numbers with the final model.
  5. [Eq. (5)-(6)] The reward definition is ambiguous and under-specified. The text says 'Rf are decomposed into two items Rtvg and Rpir', but Eq. (4) defines Rf as a format reward and Eq. (5) defines Rp = Rtvg + Rpir; the notation '||·||norm' is undefined, the IoU term in Eq. (5) does not specify how empty predictions or empty ground-truth intervals are handled, and the weights α, β, and γ in Eqs. (5)-(6) are never given numerical values. Please rewrite the reward formulation consistently and report all hyperparameter values so that the RL procedure is reproducible.
minor comments (6)
  1. [Table 2] The row labeled 'Tempo-R1' should likely read 'Time-R1', and the corresponding reference (Wang et al. 2025) should be cited in the table caption; as printed, the row appears to be a typo that obscures the comparison.
  2. [Table 7] The header 'Irrelevance Paris Ratio' should read 'Irrelevance Pairs Ratio', and the column labeled 'F Dataset R' with values 0.1, 86.6, 81.7, 94.6, and 85.5 is unexplained; please define the metric and the data split.
  3. [Explicit Timestamp Alignment (ETA)] The sentence referring to 'Figure N' contains a placeholder figure number; please replace it with the actual figure reference.
  4. [Implementation Details] Several hyperparameters are not reported: the SAA threshold T_Key, the resolution scales Rl and Rs, the reward weights α, β, and γ, the OOD content-similarity threshold, and the visual token budgets. Please include these values for reproducibility.
  5. [Testbench with More Reasonable Ground Truth (GT) Annotations] The sentence 'In subsequent work, we will use it together with the original QvHighlights to evaluate the TVG capabilities of various approaches' suggests the corrected testbench is not yet finalized; please clarify whether cQvH is a released artifact of this paper or a preview for future work.
  6. [Eq. (5)] Eq. (5) uses set operations on timestamp intervals (P rdi ∩ GTi and P rdi ∪ GTi) without defining interval intersection and union; please define the interval IoU computation explicitly.

Circularity Check

2 steps flagged · score 6.0 of 10

Partial circularity: the cQvH benchmark is author-corrected after observing the evaluated model, and the 10% irrelevance ratio is selected on the QvH test split before reporting final scores.

  1. self definitional [Section 'Testbench with More Reasonable Ground Truth (GT) Annotations'; contribution 'A More Reasonable Testbench']
    "Furthermore, during our testing process, we observe and correct some ground-truth annotations inconsistent with common sense in the QvHighlights evaluation data set. We will present the performance of Tempo-R0 on the QvHighlights dataset before and after correction in the following experiments respectively. ... To mitigate this issue, we manually rectify these erroneous annotations and provide a corrected testbench dubbed as cQvH."

    The corrected QvHighlights (cQvH) labels are generated by the authors themselves while testing the same model that is later evaluated on them. There is no blind protocol, no independent annotator agreement, and no pre-specified correction rule. The claim that cQvH is 'more consistent with human perception' is therefore self-attested by the same group whose model scores higher on those corrected labels. Evaluating Tempo-R0 on cQvH and then citing the resulting advantage is a self-referential validation loop: the reference target was constructed by observing the model's behavior, so the high cQvH scores are partly an artifact of the authors' own corrections rather than an independent benchmark.

  2. fitted input called prediction [Ablation Experiments, Table 7 'Irrelevance Ratio Comparison'; Implementation Details]
    "we select QvH test dataset to perform these ablation experiments. ... we list several results under different ratios of irrelevant video-query pairs in the entire training set and the comparison in Table 7 shows that around 10% ratio yields the averagely best results. ... In the second stage, we only use the ID training sets mixed with about ten-percent of augmented irrelevant video query training pairs."

    The 10% irrelevance ratio is chosen by comparing Table 7 scores computed on the QvH test split, and the final Tempo-R0 is then trained with 'about ten-percent' and its Table 1 QvH/cQvH scores are reported as results. This fits a key training hyperparameter to the test benchmark and then reports performance on that same benchmark. Consequently, the headline QvH advantage is not an independent out-of-sample estimate but a test-set-selected outcome, making the reported 'prediction' statistically forced by the model-selection procedure.

full rationale

Most of the pipeline is not circular in the derivation sense: SAA and ETA are preprocessing choices, and PIR-GRPO is a standard RLVR reward with IoU-based verifiable rewards, not a quantity that reduces to its own inputs. There are no load-bearing self-citations. The circularity is concentrated in the evaluation loop. First, the cQvH testbench is created by the authors after observing the model during testing, and the same model is then validated against it, so the cQvH advantage is partly self-confirmatory. Second, the 10% irrelevance ratio and other design choices are selected using ablations on the QvH test set, which compromises the independence of the final QvH and cQvH scores. The missing Qwen2-VL-7B TVG baseline in Table 1, the apparent 'Tempo-R1' typo in Table 2, and the comparison to older BLIP-2 baselines are correctness and comparability concerns rather than circularity, and do not by themselves raise the circularity score. Because the method components still have independent content on the same backbone and the original QvH benchmark is external, the paper is only partially circular, not wholly reducible to its inputs.

Assumptions & free parameters 5 free parameters · 6 assumptions · 1 invented entities

The central claim rests on several unquantified choices (SAA threshold and scales, reward weights, irrelevance ratio) and on the authors' own corrected testbench. The pretrained Qwen2-VL-7B backbone is a major external input, and the RLVR training assumes the reward functions are reliable task signals.

free parameters (5)
  • SAA threshold T_Key = not reported
    Preset threshold that decides whether a frame is downsampled to high or low resolution; no value or tuning procedure given.
  • SAA resolution scales Rl, Rs = not reported (Rl > Rs)
    Downsampling factors controlling token allocation; values not disclosed.
  • Irrelevant pair ratio = 10%
    Chosen from ablation Table 7 as the best of 0, 5, 10, 20, 30%; selection on test set is a free parameter.
  • Reward weights alpha, beta, gamma = not reported
    Weights in Eq. (6) and refusal term Eq. (5); no values given.
  • OOD data amount/content similarity threshold = not reported
    Amount of InternVid OOD data and similarity criterion used to select it are unquantified.
assumptions (6)
  • domain assumption OT-based hue-channel frame variation is a valid proxy for event-boundary salience
    Used to allocate visual tokens in SAA (Eq. 2); if false, SAA could reduce grounding accuracy despite the small ablation gain.
  • domain assumption Zero-padded textual timestamps interleaved with visual tokens improve temporal alignment in Qwen2-VL-7B
    ETA design assumes the MLLM can exploit explicit timestamps as a third modality.
  • domain assumption GRPO with a refusal reward generalizes to improve relevant-pair grounding
    PIR-GRPO assumes the RL signal on irrelevant pairs does not harm and helps acceptance; ablation does not isolate this from plain RL.
  • ad hoc to paper The manually corrected QVHighlights annotations are indeed more reasonable than the original
    The corrected testbench is produced by the authors with no inter-annotator agreement and no blinded protocol.
  • domain assumption Qwen2-VL-7B is an appropriate frozen-configuration backbone that can be fine-tuned for TVG
    The whole pipeline inherits Qwen2-VL-7B's pretrained visual and text encoders.
  • domain assumption IoU-based precision and format rewards are verifiable and sufficient training signals
    RLVR assumes the reward functions correctly measure task success.
invented entities (1)
  • Corrected QVHighlights testbench (cQvH)
    purpose: Evaluation benchmark with manually corrected ground-truth annotations
    The paper introduces a manually corrected version of the QVHighlights test set with no external validation or inter-annotator statistics; the correction process is not shown to be model-blind.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Tempo-R0: A Video-MLLM for Temporal Video Grounding through Efficient Temporal Sensing Reinforcement Learning." pith.science (2026). https://pith.science/paper/CGRJCBEZ

@misc{pith2026250704702,
  author       = {Pith},
  title        = {Pith review of: Tempo-R0: A Video-MLLM for Temporal Video Grounding through Efficient Temporal Sensing Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CGRJCBEZ}},
  note         = {Machine review of arXiv:2507.04702}
}
read the original abstract

Temporal Video Grounding (TVG), which requires pinpointing relevant temporal segments from video based on language query, has always been a highly challenging task in the field of video understanding. Videos often have a larger volume of information and redundancy than texts or images. Models should present comprehensive understanding of the whole video to accurately retrieve query-relevant clips. We thus propose Tempo-R0: a Video Multimodal Large Language Model (Video-MLLM) for the temporal video grounding task via multimodal temporal sensing reinforcement. Specifically, during the preprocessing stage of our pipeline, we employ Self-adaptive Attention Allocation (SAA) method based on frame content variation to efficiently use the MLLM's limited attention. The Explicit Timestamp-modal Aligned (ETA) method is also utilized to strengthen our model's capability to perceive the boundaries of events in the video. In the fine-tuning part of our pipeline, we creatively apply Partial Irrelevance Refusing-based Group Relative Policy Optimization (PIR-GRPO) in TVG area to foster model's temporal reasoning from not only accepting relevant video-query pairs but also refusing irrelevant ones. Experiments demonstrate that our method accomplishes a notable advantage over SOTA solutions by around 3.5% on both the original QVHighlights testbench and its corrected version with more reasonable ground truth annotations.

Figures

Figures reproduced from arXiv: 2507.04702 by the authors.

Figure 1
Figure 1. The major improvements realized by our Tempo-R0 compared to other Video-MLLMs. adding pure irrelevant videos training data containing no relevant information to queries and corresponding adaption on the original GRPO method, we teach the MLLM to learn effective refusion and furthermore im￾prove MLLM’s average performance for TVG tasks in our two-staged RFT fine-tuning process. • A More Reasonable Testbench: Through … view at source ↗
Figure 2
Figure 2. In the inference pipeline of Tempo-R0, the inputs are decomposed into three independent modalities for separate [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The advantage of OT-based methods in key image [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: One of the inappropriate GT annotation example in the QvHighlights testbench with our corrected annotation. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: The comparison of a VTG example in QvH and cQvH testbenchs. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. Video-OPD: Efficient Post-Training of Multimodal Large Language Models for Temporal Video Grounding via On-Policy Distillation

    cs.CV 2026-02 unverdicted novelty 7.0 of 10

    Video-OPD uses on-policy distillation from a frontier teacher to turn sparse episode rewards into dense step-wise signals for more efficient post-training of MLLMs on temporal video grounding.

  2. TimePLE: Rethinking Temporal Representation for Video Temporal Grounding

    cs.CV 2026-07 conditional novelty 6.0 of 10

    TimePLE predicts a whole video interval as a joint distribution over a position-duration square, rather than predicting start and end separately, and reports higher mIoU across four VTG benchmarks.

Reference graph

Works this paper leans on

22 extracted references · 8 canonical work pages · cited by 2 Pith papers

  1. [3]

    arXiv:2505.09568

    BLIP3-o: A Family of Fully Open Uni- fied Multimodal Models-Architecture, Training and Dataset. arXiv:2505.09568. Feichtenhofer, C.; Fan, H.; Malik, J.; and He, K

    arXiv

  2. [4]

    anthropic.com/news/claude-4

    https://www. anthropic.com/news/claude-4. Accessed: 2025-07-02. Bai, S.; Chen, K.; Liu, X.; Wang, J.; Ge, W.; Song, S.; Dang, K.; Wang, P.; Wang, S.; Tang, J.; Zhong, H.; Zhu, Y .; Yang, M.; Li, Z.; Wan, J.; Wang, P.; Ding, W.; Fu, Z.; Xu, Y .; Ye, J.; Zhang, X.; Xie, T.; Cheng, Z.; Zhang, H.; Yang, Z.; Xu, H.; and Lin, J

    work page 2025

  3. [5]

    In 2025 IEEE/CVF Winter Conference on Appli- cations of Computer Vision (WACV), 5336–5345

    Moment of Untruth: Dealing with Negative Queries in Video Moment Retrieval. In 2025 IEEE/CVF Winter Conference on Appli- cations of Computer Vision (WACV), 5336–5345. IEEE. Gao, J.; Sun, C.; Yang, Z.; and Nevatia, R

    work page 2025

  4. [11]

    arXiv:2301.12597

    BLIP-2: Boot- strapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models. arXiv:2301.12597. Li, J.; Li, D.; Xiong, C.; and Hoi, S

    arXiv

  5. [13]

    arXiv:2504.06958

    VideoChat-R1: Enhancing Spatio-Temporal Perception via Reinforcement Fine-Tuning. arXiv:2504.06958. Liu, H.; Li, C.; Li, Y .; and Lee, Y . J

    arXiv

  6. [14]

    arXiv:2310.03744

    Improved Base- lines with Visual Instruction Tuning. arXiv:2310.03744. Meinardus, B.; Rodriguez, H.; Batra, A.; Rohrbach, A.; and Rohrbach, M

    arXiv

  7. [15]

    arXiv:2406.18113

    Chrono: A Simple Blueprint for Repre- senting Time in MLLMs. arXiv:2406.18113. Moon, W.; Hyun, S.; Lee, S.; and Heo, J.-P. 2023a. Correlation-Guided Query-Dependency Calibration for Video Temporal Grounding. arXiv:2311.08835. Moon, W.; Hyun, S.; Park, S.; Park, D.; and Heo, J.-P. 2023b. Query-dependent video representation for moment retrieval and highlig...

    arXiv

  8. [16]

    https://openai

    OpenAI o1 System Card. https://openai. com/index/openai-o1-system-card/. Accessed: 2025-07-02. OpenAI

    work page 2025

Show all 22 references
  1. [17]

    https:// openai.com/index/introducing-o3-and-o4-mini/

    Introducing OpenAI o3 and o4-mini. https:// openai.com/index/introducing-o3-and-o4-mini/. Accessed: 2025-07-02. Peyr´e, G

    2025

  2. [18]

    arXiv:2505.06589

    Optimal Transport for Machine Learners. arXiv:2505.06589. Wang, P.; Bai, S.; Tan, S.; Wang, S.; Fan, Z.; Bai, J.; Chen, K.; Liu, X.; Wang, J.; Ge, W.; Fan, Y .; Dang, K.; Du, M.; Ren, X.; Men, R.; Liu, D.; Zhou, C.; Zhou, J.; and Lin, J. 2024a. Qwen2-VL: Enhancing Vision- Lang...

  3. [19]

    arXiv:2212.03191

    InternVideo: General Video Foundation Models via Generative and Dis- criminative Learning. arXiv:2212.03191. Wang, Y .; Meng, X.; Liang, J.; Wang, Y .; Liu, Q.; and Zhao, D. 2024c. HawkEye: Training Video-Text LLMs for Grounding Text in Videos. arXiv:2403.10228. Wang, Y .; Wan...

    arXiv

  4. [20]

    arXiv:2503.13377

    Time-R1: Post-Training Large Vision Language Model for Temporal Video Grounding. arXiv:2503.13377. Xue, L.; Shu, M.; Awadalla, A.; Wang, J.; Yan, A.; Purush- walkam, S.; Zhou, H.; Prabhu, V .; Dai, Y .; Ryoo, M. S.; Kendre, S.; Zhang, J.; Tseng, S.; Lujan-Moreno, G. A.; Olson,...

    arXiv

  5. [21]

    arXiv:2408.08872

    xGen-MM (BLIP-3): A Family of Open Large Multimodal Models. arXiv:2408.08872. Yu, S.; Cho, J.; Yadav, P.; and Bansal, M

  6. [22]

    arXiv:2305.06988

    Self- Chained Image-Language Model for Video Localization and Question Answering. arXiv:2305.06988

    arXiv

  7. [2015]

    In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 961–

    ActivityNet: A large-scale video benchmark for hu- man activity understanding. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 961–

    2015

  8. [2017]

    In 2017 IEEE International Conference on Computer Vision (ICCV), 5277–5285

    TALL: Temporal Activity Localization via Language Query. In 2017 IEEE International Conference on Computer Vision (ICCV), 5277–5285. IEEE. Gordeev, A.; Dokholyan, V .; Tolstykh, I.; and Kuprashevich, M

    2017

  9. [2019]

    In 2019 IEEE/CVF International Conference on Computer Vision (ICCV)

    SlowFast Networks for Video Recognition. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE. Flanagan, K.; Damen, D.; and Wray, M

    2019

  10. [2021]

    arXiv:2107.09609

    QVHighlights: De- tecting Moments and Highlights in Videos via Natural Lan- guage Queries. arXiv:2107.09609. Li, J.; Li, D.; Savarese, S.; and Hoi, S

    arXiv

  11. [2022]

    arXiv:2201.12086

    BLIP: Bootstrapping Language-Image Pre-training for Uni- fied Vision-Language Understanding and Generation. arXiv:2201.12086. Li, X.; Yan, Z.; Meng, D.; Dong, L.; Zeng, X.; He, Y .; Wang, Y .; Qiao, Y .; Wang, Y .; and Wang, L

    arXiv

  12. [2023]

    In 2023 IEEE/CVF International Con- ference on Computer Vision (ICCV), 13800–13810

    Knowing Where to Focus: Event-aware Transformer for Video Grounding. In 2023 IEEE/CVF International Con- ference on Computer Vision (ICCV), 13800–13810. IEEE. Lei, J.; Berg, T. L.; and Bansal, M

    2023

  13. [2024]

    arXiv:2410.01615

    Saliency-Guided DETR for Moment Retrieval and Highlight Detection. arXiv:2410.01615. Heilbron, F. C.; Escorcia, V .; Ghanem, B.; and Niebles, J. C

  14. [2025]

    arXiv:2502.13923

    Qwen2.5-VL Technical Report. arXiv:2502.13923. Chen, J.; Xu, Z.; Pan, X.; Hu, Y .; Qin, C.; Goldstein, T.; Huang, L.; Zhou, T.; Xie, S.; Savarese, S.; Xue, L.; Xiong, C.; and Xu, R

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.