Pith. sign in

REVIEW 4 major objections 2 minor 39 references

Aligning Moments in Time using Video Queries

T0 review · 4 major / 2 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A video-query transformer claims state-of-the-art moment retrieval, but the manuscript body is an unrelated pulsar study.

desk verdict Abstract claims SOTA video moment retrieval; the body is an unrelated radio pulsar paper — nothing in the submission supports the abstract. read the letter →

arxiv 2508.15439 v2 pith:ALHHKS3L submitted 2025-08-21 cs.CV

Yogesh Kumar , Uday Agarwal , Manish Gupta , Anand Mishra This is my paper · ORCID
classification cs.CV
keywords video-to-videomomentretrievalAlignmentTransformerMATRdual-stagesequenceself-supervisedpretrainingActivityNet-VRLSportsMoments
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 abstract claims that a new architecture called MATR (Moment Alignment Transformer) solves video-to-video moment retrieval: given a query video clip, it finds the semantically matching segment in a target video, and it reports large gains over prior methods on two benchmarks. If true, this would let users search video by example rather than by text. The accompanying manuscript, however, is a radio-astronomy paper titled 'Radio Observations of a Candidate Redback Millisecond Pulsar: 1FGL J0523.5−2529'; it contains no architecture, experiments, or dataset related to MATR. Read in good faith, the abstract's claim is unverifiable from the provided text, and the load-bearing premise — that the body belongs to the abstract — fails on inspection.

What carries the argument

MATR — the Moment Alignment Transformer. The load-bearing components are dual-stage sequence alignment, which conditions target-video features on query-video features to capture cross-video dependencies, and a self-supervised pretraining task where the model localizes random clips inside videos, followed by foreground/background classification and boundary-prediction heads. In the supplied full text none of these components appear; they exist only in the abstract's description.

What would settle it

Inspect the full text of the submission: it is titled 'Radio Observations of a Candidate Redback Millisecond Pulsar: 1FGL J0523.5−2529', contains telescope observation tables, and never mentions MATR, dual-stage sequence alignment, ActivityNet-VRL, or SportsMoments. No further computation is needed to confirm that the abstract's claims are unsupported by the provided document.

Watch

Extended reading notes

Core claim

The paper's stated discovery is that MATR, a transformer that conditions target-video representations on query-video features through dual-stage sequence alignment, improves moment localization by 13.1% in R@1 and 8.1% in mIoU absolute over state-of-the-art on ActivityNet-VRL, and 14.7% in R@1 and 14.4% in mIoU on the new SportsMoments dataset, using a self-supervised clip-localization pretraining. The full-text document provided for this submission, however, is an unrelated observational study of a candidate redback millisecond pulsar; it does not describe MATR, the alignment mechanism, the pretraining, either dataset, or any of the reported numbers. The author's intended assertion is there

Load-bearing premise

The central claim depends on the manuscript body being the paper described by the abstract, but the body is an unrelated radio-astronomy study titled 'Radio Observations of a Candidate Redback Millisecond Pulsar: 1FGL J0523.5−2529' — that premise fails on inspection.

Editorial extensions

If this is right

  • Video-to-video moment retrieval would become practical: users could query by example clip rather than by text, capturing actions words cannot describe.
  • Self-supervised clip-localization pretraining would transfer to a harder cross-video localization task, reducing the need for labeled video-moment data.
  • The reported gains — 13.1% R@1 and 8.1% mIoU on ActivityNet-VRL; 14.7% R@1 and 14.4% mIoU on SportsMoments — would set a new state of the art on both benchmarks.
  • The new SportsMoments dataset would provide a sports-specific benchmark for video-query retrieval, complementing ActivityNet-VRL.

Reading between the lines

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

  • The abstract and the body appear to come from different submissions; if so, the abstract's experimental numbers can only be taken as unsupported until the matching manuscript is located and audited.
  • Independently of this mismatched document, the claim that self-supervised clip localization transfers to video-to-video moment retrieval is a testable hypothesis: one could pretrain a transformer on clip localization, fine-tune on a labeled video-retrieval benchmark, and measure transfer on held-out domains.
  • A reader seeking the method would need to find the actual MATR paper separately; the currently supplied full text provides no implementation details, architecture diagram, or hyperparameter settings to reproduce.
  • The reported performance gaps over prior art are large enough that, if verified, they would likely drive adoption of dual-stage cross-video conditioning in other video-understanding tasks such as action localization and dense video captioning.
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 / 2 minor

Summary. The submission claims to introduce MATR, a transformer-based model for video-to-video moment retrieval, with abstract-level results of 13.1% R@1 and 8.1% mIoU absolute improvement over state-of-the-art on ActivityNet-VRL and 14.7% R@1 / 14.4% mIoU on a new SportsMoments dataset. The full text, however, is an unrelated radio-astronomy manuscript, 'Radio Observations of a Candidate Redback Millisecond Pulsar: 1FGL J0523.5-2529', whose author list, abstract, Introduction, Methods, Results, and references concern pulsar timing observations. No MATR architecture, no dual-stage sequence alignment, no pre-training objective, no dataset description, no experimental protocol, and no results tables appear anywhere in the body.

Significance. If the abstract claims were accompanied by the described model, experiments, and dataset, MATR would represent a substantial advance in video-to-video moment retrieval. The claimed absolute gains of 13.1% R@1 and 8.1% mIoU over prior methods are large, and a new SportsMoments dataset could be a useful community resource. However, none of these contributions are present in the submitted manuscript. The body contains no method to audit, no code or machine-checked artifacts, and no quantitative evaluation relating to moment retrieval. At the document level, the paper reduces to an abstract-only claim with an unrelated supporting text. The significance cannot be assessed because the contribution itself is absent.

major comments (4)
  1. [Abstract vs. full text] The central claims of the paper, namely the MATR architecture, dual-stage sequence alignment, self-supervised pre-training, and the ActivityNet-VRL and SportsMoments results, appear only in the abstract. The full text is an astronomy paper about radio observations of the candidate redback pulsar 1FGL J0523.5-2529, as confirmed by the title, author list, abstract, Sections 1-5, references, and the footer 'arXiv:2508.15435v2 [astro-ph.HE]'. There is no moment-retrieval algorithm, no model diagram, no equations describing attention or alignment, and no dataset. The submitted manuscript is therefore not a paper containing the claimed contribution.
  2. [§4 and Table 3] The only quantitative results in the body are flux-density upper limits for a pulsar non-detection (Table 3) and related detection limits in Eqs. (3)-(4). These are unrelated to the abstract's R@1 and mIoU claims. The 13.1% R@1, 8.1% mIoU, 14.7% R@1, and 14.4% mIoU figures are not supported by any experimental protocol, baseline description, error bars, or comparison table in the manuscript. There is no way to verify, reproduce, or even locate the source of these numbers.
  3. [Dataset and pre-training] The abstract announces 'our newly proposed dataset, SportsMoments' and a self-supervised clip-localization pre-training technique. Neither is described in the manuscript. The dataset name does not occur in the body, and no annotation statistics, collection procedure, evaluation splits, or baseline results are given. The claim that self-supervised clip localization transfers to video-to-video moment retrieval is thus a bare assertion; no experiments address it.
  4. [Document integrity] The manuscript's own footer identifies the body as arXiv:2508.15435v2 [astro-ph.HE], an arXiv ID different from the submission ID 2508.15439. Internal cross-references (e.g., Section 2 observations, Figure 1) all refer to telescope observations, not to video retrieval. This is not a presentation issue or a missing appendix; it is an absence of the paper being reviewed.
minor comments (2)
  1. [Metadata] The arXiv ID and subject class on the footer do not match the submission's stated ID and category; this should be corrected by resubmitting the intended paper.
  2. [Throughout] The astronomy body has several typographical issues (e.g., 'Trinty' for 'Trinity', 'spliting', 'P b≤10' formatting), but these are irrelevant to the video-retrieval manuscript and are listed only for completeness.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation identified: the supplied body is an unrelated astronomy paper, so there is no method chain to reduce.

full rationale

The abstract describes a video-to-video moment retrieval model (MATR) with transformer architecture, dual-stage sequence alignment, self-supervised pre-training, and quantitative SOTA gains on ActivityNet-VRL and a new SportsMoments dataset. The full text supplied, however, is an unrelated astrophysics manuscript titled 'RADIO OBSERVATIONS OF A CANDIDATE REDBACK MILLISECOND PULSAR: 1FGL J0523.5−2529' (footer arXiv:2508.15435v2 [astro-ph.HE]). It contains no equations for MATR, no alignment module, no pre-training objective, no dataset description, and no results tables. There is therefore no claimed derivation chain that can be walked, and I cannot exhibit any specific reduction in which a prediction is equivalent to its input by construction, a fitted parameter is renamed as a prediction, or a load-bearing conclusion rests solely on a self-citation. The document-level mismatch is a severe absence of supporting content and a correctness/integrity problem, but it is not circularity under the definitions used here. The secondary assumption that self-supervised clip-localization pre-training transfers to video-to-video moment retrieval is untested, but an untested transfer assumption is not a circular step. Score 0 reflects the absence of any identifiable circular step; it should not be read as validating the abstract's claims.

Assumptions & free parameters 0 free parameters · 3 assumptions · 1 invented entities

The ledger is necessarily minimal because the manuscript body does not contain the model, its hyperparameters, or its training details. No free parameters can be named from the abstract. The assumptions listed are the domain assumptions the abstract's framing rests on, and they inherit the document-level mismatch. The only introduced artifact named in the abstract is the SportsMoments dataset, which has no release or external handle in the text. The unrelated body text introduces astrophysical quantities (dispersion measure, Z-parameter, flux limits), but those bear on a different claim and are not counted here.

assumptions (3)
  • domain assumption Semantic frame-level alignment between query and target video frames is a sufficient and learnable signal for precise moment localization.
    The abstract frames the task around 'semantic frame-level alignment'; the entire MATR design assumes this signal exists and can be learned from paired video data. No evidence for the assumption appears in the manuscript body.
  • domain assumption Self-supervised pretraining, localizing random clips within videos, transfers to the video-to-video moment retrieval task.
    The abstract proposes this initialization as a strength; it assumes the proxy task shares structure with the downstream task. The premise is not evidenced in the body, which does not describe the pretraining at all.
  • domain assumption The ActivityNet-VRL split and the constructed SportsMoments dataset correctly instantiate the video-to-video retrieval task.
    All reported gains are measured on these benchmarks, so their annotation semantics and splits are load-bearing. The body gives no dataset statistics or annotation protocol for SportsMoments and no details for ActivityNet-VRL usage.
invented entities (1)
  • SportsMoments dataset
    purpose: New benchmark for video-to-video moment retrieval, introduced to evaluate MATR.
    The abstract introduces this dataset but gives no release location, statistics, or download link, so it provides no falsifiable handle outside the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Aligning Moments in Time using Video Queries." pith.science (2026). https://pith.science/paper/ALHHKS3L

@misc{pith2026250815439,
  author       = {Pith},
  title        = {Pith review of: Aligning Moments in Time using Video Queries},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ALHHKS3L}},
  note         = {Machine review of arXiv:2508.15439}
}
read the original abstract

Video-to-video moment retrieval (Vid2VidMR) is the task of localizing unseen events or moments in a target video using a query video. This task poses several challenges, such as the need for semantic frame-level alignment and modeling complex dependencies between query and target videos. To tackle this challenging problem, we introduce MATR (Moment Alignment TRansformer), a transformer-based model designed to capture semantic context as well as the temporal details necessary for precise moment localization. MATR conditions target video representations on query video features using dual-stage sequence alignment that encodes the required correlations and dependencies. These representations are then used to guide foreground/background classification and boundary prediction heads, enabling the model to accurately identify moments in the target video that semantically match with the query video. Additionally, to provide a strong task-specific initialization for MATR, we propose a self-supervised pre-training technique that involves training the model to localize random clips within videos. Extensive experiments demonstrate that MATR achieves notable performance improvements of 13.1% in R@1 and 8.1% in mIoU on an absolute scale compared to state-of-the-art methods on the popular ActivityNet-VRL dataset. Additionally, on our newly proposed dataset, SportsMoments, MATR shows a 14.7% gain in R@1 and a 14.4% gain in mIoU on an absolute scale over strong baselines.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 14 canonical work pages

  1. [1]

    A., Ajello, M., Allafort, A., et al

    Abdo, A. A., Ajello, M., Allafort, A., et al. 2013, The Astrophysical Journal Supplement Series, 208, 17, doi:10.1088/0067-0049/208/2/17

    doi:10.1088/0067-0049/208/2/17 2013

  2. [2]

    2012, The Astrophysical Journal, 753, 83, doi:10.1088/0004-637X/753/1/83

    Ackermann, M., Ajello, M., Allafort, A., et al. 2012, The Astrophysical Journal, 753, 83, doi:10.1088/0004-637X/753/1/83

    doi:10.1088/0004-637x/753/1/83 2012

  3. [3]

    D., Thornton, D., Bailes, M., et al

    Bates, S. D., Thornton, D., Bailes, M., et al. 2015, MNRAS, 446, 4019, doi:10.1093/mnras/stu2350

    work page doi:10.1093/mnras/stu2350 2015

  4. [4]

    Bhattacharya, D., & van den Heuvel, E. P. J. 1991, Phys. Rep., 203, 1, doi:10.1016/0370-1573(91)90064-S

    doi:10.1016/0370-1573(91)90064-s 1991

  5. [5]

    W., Fender, R

    Broderick, J. W., Fender, R. P., Breton, R. P., et al. 2016, Monthly Notices of the Royal Astronomical Society, 459, 2681, doi:10.1093/mnras/stw794

    work page doi:10.1093/mnras/stw794 2016

  6. [6]

    M., & Han, Z

    Chen, H.-L., Chen, X., Tauris, T. M., & Han, Z. 2013, ApJ, 775, 27, doi:10.1088/0004-637X/775/1/27

    doi:10.1088/0004-637x/775/1/27 2013

  7. [7]

    M., & Lazio, T

    Cordes, J. M., & Lazio, T. J. W. 2003, NE2001.I. A New Model for the Galactic Distribution of Free Electrons and its Fluctuations, arXiv, doi:10.48550/arXiv.astro-ph/0207156 De Vito, M. A., Benvenuto, O. G., & Horvath, J. E. 2020, MNRAS, 493, 2171, doi:10.1093/mnras/staa395

    arXiv doi:10.48550/arxiv.astro-ph/0207156 2003

  8. [8]

    S., Ray, P

    Deneva, J. S., Ray, P. S., Camilo, F., et al. 2016, ApJ, 823, 105, doi:10.3847/0004-637X/823/2/105

    work page doi:10.3847/0004-637x/823/2/105 2016

Show all 39 references
  1. [9]

    S., & Backer, D

    Foster, R. S., & Backer, D. C. 1990, ApJ, 361, 300, doi:10.1086/169195

    1990 doi

  2. [10]

    P., Perez, K

    Halpern, J. P., Perez, K. I., & Bogdanov, S. 2022, The Astrophysical Journal, 935, 151, doi:10.3847/1538-4357/ac8161

    2022 doi

  3. [11]

    N., Dunning, A., et al

    Hobbs, G., Manchester, R. N., Dunning, A., et al. 2020, Publications of the Astronomical Society of Australia, 37, e012, doi:10.1017/pasa.2020.2

    2020 doi

  4. [12]

    F., et al

    Jankowski, F., van Straten, W., Keane, E. F., et al. 2018, MNRAS, 473, 4436, doi:10.1093/mnras/stx2476

    2018 doi

  5. [13]

    F., Barr, E

    Keane, E. F., Barr, E. D., Jameson, A., et al. 2018, Monthly Notices of the Royal Astronomical Society, 473, 116, doi:10.1093/mnras/stx2126

    2018 doi

  6. [14]

    Y ., Zharikov, S

    Kirichenko, A. Y ., Zharikov, S. V ., Karpova, A. V ., et al. 2024, Monthly Notices of the Royal Astronomical Society, 527, 4563, doi:10.1093/mnras/stad3391

    2024 doi

  7. [15]

    Koljonen, K. I. I., & Linares, M. 2025, SpiderCat: A Catalog of Compact Binary Millisecond Pulsars, arXiv, doi:10.48550/arXiv.2505.11691

    2025 doi

  8. [16]

    H., Manchester, R

    Kramer, M., Stairs, I. H., Manchester, R. N., et al. 2021, Physical Review X, 11, 041050, doi:10.1103/PhysRevX.11.041050

    2021 doi

  9. [17]

    R., & Kramer, M

    Lorimer, D. R., & Kramer, M. 2005, Handbook of Pulsar Astronomy, Cambridge Observing Handbooks for Research Astronomers (Cambridge University Press)

    2005

  10. [18]

    G., Manchester, R

    Lyne, A. G., Manchester, R. N., Lorimer, D. R., et al. 1998, MNRAS, 295, 743, doi:10.1046/j.1365-8711.1998.01144.x

    1998

  11. [19]

    N., Hobbs, G

    Manchester, R. N., Hobbs, G. B., Teoh, A., & Hobbs, M. 2005, The Astronomical Journal, 129, 1993, doi:10.1086/428488

    2005 doi

  12. [20]

    2024, Astronomy and Astrophysics, 683, A183, doi:10.1051/0004-6361/202348247

    Men, Y ., & Barr, E. 2024, Astronomy and Astrophysics, 683, A183, doi:10.1051/0004-6361/202348247

    2024 doi

  13. [21]

    2022, in Astrophysics and Space Science

    Papitto, A., & de Martino, D. 2022, in Astrophysics and Space Science

    2022

  14. [22]

    465, Physics and Astrophysics of Neutron Stars (Springer Nature), 157–200, doi:10.1007/978-3-030-85198-9_6

    Library, V ol. 465, Physics and Astrophysics of Neutron Stars (Springer Nature), 157–200, doi:10.1007/978-3-030-85198-9_6

    doi

  15. [23]

    2013, Nature, 501, 517, doi:10.1038/nature12470

    Papitto, A., Ferrigno, C., Bozzo, E., et al. 2013, Nature, 501, 517, doi:10.1038/nature12470

    2013 doi

  16. [24]

    2011, Astrophysics Source Code Library, ascl:1107.017

    Ransom, S. 2011, Astrophysics Source Code Library, ascl:1107.017. https://ui.adsabs.harvard.edu/abs/2011ascl.soft07017R

    2011

  17. [25]

    M., Greenhill, L

    Ransom, S. M., Greenhill, L. J., Herrnstein, J. R., et al. 2001, ApJ, 546, L25, doi:10.1086/318062

    2001 doi

  18. [26]

    Roberts, M. S. E., Noori, H. A., Torres, R. A., et al. 2017, Proceedings of the International Astronomical Union, 13, 43, doi:10.1017/S1743921318000480

    2017 doi

  19. [27]

    S., Bhattacharyya, B., et al

    Roy, J., Ray, P. S., Bhattacharyya, B., et al. 2015, The Astrophysical Journal Letters, 800, L12, doi:10.1088/2041-8205/800/1/L12

    2015 doi

  20. [28]

    2016, PhD thesis, University of Virginia

    Sanpa-Arsa, S. 2016, PhD thesis, University of Virginia

    2016

  21. [29]

    L., Tout, C

    Smedley, S. L., Tout, C. A., Ferrario, L., & Wickramasinghe, D. T. 2015, Monthly Notices of the Royal Astronomical Society, 446, 2540, doi:10.1093/mnras/stu2252

    2015 doi

  22. [30]

    W., Archibald, A

    Stappers, B. W., Archibald, A. M., Hessels, J. W. T., et al. 2014, The Astrophysical Journal, 790, 39, doi:10.1088/0004-637X/790/1/39

    2014 doi

  23. [31]

    2016, ApJ, 833, 192, doi:10.3847/1538-4357/833/2/192

    Stovall, K., Allen, B., Bogdanov, S., et al. 2016, ApJ, 833, 192, doi:10.3847/1538-4357/833/2/192

    2016 doi

  24. [32]

    2014, The Astrophysical Journal, 788, L27, doi:10.1088/2041-8205/788/2/L27

    Strader, J., Chomiuk, L., Sonbas, E., et al. 2014, The Astrophysical Journal, 788, L27, doi:10.1088/2041-8205/788/2/L27

    2014 doi

  25. [33]

    M., & van den Heuvel, E

    Tauris, T. M., & van den Heuvel, E. P. J. 2006, in Compact stellar X-ray sources, ed. W. H. G. Lewin & M. van der Klis, V ol. 39, 623–665, doi:10.48550/arXiv.astro-ph/0303456

    2006 doi arXiv

  26. [34]

    J., Breton, R

    Thongmeearkom, T., Clark, C. J., Breton, R. P., et al. 2024, Monthly Notices of the Royal Astronomical Society, 530, 4676, doi:10.1093/mnras/stae787

    2024 doi

  27. [35]

    M., Manchester, R

    Yao, J. M., Manchester, R. N., & Wang, N. 2017, The Astrophysical Journal, 835, 29, doi:10.3847/1538-4357/835/1/29

    2017 doi

  28. [36]

    M., D’Avanzo, P., Ridolfi, A., et al

    Zanon, A. M., D’Avanzo, P., Ridolfi, A., et al. 2021, Astronomy & Astrophysics, 649, A120, doi:10.1051/0004-6361/202040071

    2021 doi

  29. [37]

    S., et al

    Zheng, H., Tegmark, M., Dillon, J. S., et al. 2017, Monthly Notices of the Royal Astronomical Society, 464, 3486, doi:10.1093/mnras/stw2525

    2017 doi

  30. [38]

    2022, Research in Astronomy and Astrophysics, 22, 085001, doi:10.1088/1674-4527/ac712b

    Zhou, Z.-R., Wang, J.-B., Wang, N., Hobbs, G., & Wang, S.-Q. 2022, Research in Astronomy and Astrophysics, 22, 085001, doi:10.1088/1674-4527/ac712b

    2022 doi

  31. [39]

    Zic, A., Wang, Z., Lenc, E., et al. 2024, Monthly Notices of the Royal Astronomical Society, 528, 5730, doi:10.1093/mnras/stae033 6.APPENDIX 6.1.Detection of J0520-2553 As stated in §4 there was a serendipitous detection of J0520-2553 (see Fig. 6), using both the acceleration ...

    2024 doi

Pith tools

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