{"id":"4f06de25-11c9-41ec-ae0f-b42b77718c57","arxiv_id":"2506.03473","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MamFusion adds a Mamba module and two temporal cross-attention modules to GMMFormer, achieving slight SumR improvements on three partially relevant video retrieval benchmarks.","lead":"This paper proposes a video-retrieval model that combines a Mamba state-space module with two cross-attention fusion layers to find short relevant moments inside long, untrimmed videos. It reports small accuracy gains over the GMMFormer baseline on three public datasets, though the improvements are modest and the paper omits several experimental details.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unspecified PRVR label construction and the statement that moment annotations were 'not used' make Table I's SOTA numbers unverifiable.","rationale":"After reading the paper in good faith, the central claim is empirical state-of-the-art retrieval effectiveness (abstract and conclusion). The architecture is plausible, and no formal mathematical error is apparent. However, the evaluation is the weakest link. The paper neither gives the loss function nor specifies how video-level relevance labels are generated. The explicit sentence that moment annotations were 'not used in our experiments' is particularly concerning because the standard PRVR labels are defined by moment annotations. If the sentence is literal, the reported retrieval task is not PRVR. If it is imprecise, the missing pair-construction protocol still prevents reproduction. This concern aligns exactly with the reader's weakest_assumption, so I agree with the reader's conditional verdict. The ablation table has secondary inconsistencies (e.g., w/o Multi-Mamba SumR 69.5 does not match Table I GMMFormer 72.9 despite the text saying the Mamba block is replaced with GMMFormer), which further supports the need for code release. A direct code inspection would settle whether the evaluation protocol is standard; pending that, CONDITIONAL remains the right verdict.","tokens_in":9237,"tokens_out":6934,"duration_ms":68552,"concrete_test":"Open the repository at https://github.com/Vision-Multimodal-Lab-HZCU/MamFusion and inspect the dataset/preprocessing code that constructs (query, video) pairs for ActivityNet, Charades-STA, and TVR. Verify whether the label is derived from moment annotations (e.g., query belongs to an annotated moment of the video) and identify the loss function. Then rerun evaluation on the standard PRVR splits; if the label construction differs from MS-SL or no loss is present, Table I cannot be reproduced and the SOTA claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section III-A.1, the paper states: 'since our task focuses on video retrieval rather than moment localization, the moment annotations provided by these datasets were not used in our experiments.' In the standard PRVR protocol introduced by MS-SL [7], moment annotations define the relevance labels: a video is positive for a query if and only if the query text corresponds to one of that video's annotated moments. If those annotations were truly unused, there is no ground-truth label set for ActivityNet Captions, Charades-STA, or TVR, and the R@K/SumR values in Table I are not defined. If instead only the moment timestamps were unused at inference, the wording is misleading and the exact pair-construction procedure is still absent. Section II also contains no loss function, so the training objective is unknown. Because every reported baseline comparison inherits this label construction, the central 'state-of-the-art' claim is unverifiable unless the authors release the exact pair-generation code and loss. The claim is additionally fragile because on the primary R@1 metric MamFusion is worse than GMMFormer on ActivityNet (8.0 vs 8.3) and Charades-STA (2.0 vs 2.1); the paper does not state that SumR is the selected criterion.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MamFusion, a partially relevant video retrieval (PRVR) model built on top of GMMFormer. It introduces a Multi-Mamba module intended to handle long untrimmed video sequences, along with two temporal fusion modules (Temporal V-to-T Fusion and Temporal T-to-V Fusion) that cross-attend between text and video representations. The model is evaluated on ActivityNet Captions, Charades-STA, and TVR, and the abstract and conclusion claim state-of-the-art retrieval accuracy. The method section describes sentence and video feature extraction, GMMFormer blocks, Mamba processing, and the two fusion modules, but it does not specify a training loss or the exact protocol for constructing video-level relevance labels. The experiments report R@K and SumR metrics, with an ablation study on Charades-STA. The paper also includes a convergence speed comparison and a code link.","tokens_in":9475,"tokens_out":5708,"duration_ms":53269,"significance":"If the claims were fully substantiated, the paper would show that a selective state-space model (Mamba) can improve PRVR retrieval over the transformer-based GMMFormer, and that bidirectional temporal fusion between video and text is beneficial. The paper makes a concrete architectural contribution and provides a public code link, which are positive elements. However, the central state-of-the-art claim is currently unverifiable because the label construction process is not specified, the training loss is omitted, the reported R@1 numbers do not consistently favor MamFusion over GMMFormer, and the ablation table contradicts the narrative for the V-to-T module. The paper also omits recent PRVR baselines (DL-DKD and PEAN) that it cites in the introduction. The potential significance is real, but the present manuscript does not yet establish it rigorously.","major_comments":[{"comment":"The statement that 'moment annotations provided by these datasets were not used in our experiments' makes the ground truth for the reported R@K/SumR numbers undefined. In the standard PRVR protocol, video-level relevance labels are derived from moment annotations. Please specify exactly how positive and negative video-query pairs are constructed for each dataset; if moment annotations are used only for labeling and not at inference, state that explicitly and provide the pair-generation procedure. Without this, all results in Table I are unverifiable.","section":"Section III-A.1"},{"comment":"No training loss is specified anywhere in the method section. The paper describes the model architecture and gives equations (1)-(5), but never states the objective used for optimization (e.g., a contrastive loss over cosine similarities, with what negative sampling strategy). Add the loss formulation and training details, since the reported retrieval performance depends on the training objective.","section":"Section II"},{"comment":"The abstract and conclusion claim state-of-the-art retrieval accuracy, but on the metric declared primary in Section III-A.2 (R@K, K=1,5,10,100), MamFusion is worse than GMMFormer on R@1 on ActivityNet Captions (8.0 vs 8.3) and Charades-STA (2.0 vs 2.1). The claim is only defensible if SumR is the designated primary metric; please state this explicitly and justify it, or adjust the claim.","section":"Section III-B, Table I, abstract"},{"comment":"Table II shows that removing Temporal V-to-T Fusion ('w/o V2T') yields R@1=2.3, which is higher than the full model's R@1=2.0. The text states that removing either fusion module 'resulted in a significant drop in retrieval performance,' which is contradicted by this row. Please reconcile this discrepancy or provide a corrected ablation table and a revised interpretation of V2T's contribution.","section":"Section III-C, Table II"},{"comment":"The comparison set omits DL-DKD [18] and PEAN [19], both of which are introduced in Section I as existing PRVR methods. A state-of-the-art claim against 'existing methods' requires these recent baselines. Also report the variance or standard deviation across multiple runs for the key numbers, since the differences between models are small (e.g., SumR 177.5 vs 176.6 on TVR).","section":"Section III-B, Table I"}],"minor_comments":[{"comment":"The subsection titles and body text are swapped: Section II-C.1 is titled 'Temporal V-to-T Fusion' but describes the module as 'Temporal T-to-V Fusion,' and Section II-C.2 has the reverse error; equations (3) and (4) use TVT and TTV respectively. Please align the names.","section":"Section II-C"},{"comment":"Table II uses R@2 instead of R@5, which is inconsistent with the metric definition given in this section; please standardize.","section":"Section III-A.2"},{"comment":"The heading 'Inference Convergence Speed' describes training loss convergence; please rename the section (e.g., 'Training Convergence').","section":"Section III-D"},{"comment":"The implementation details omit the optimizer, learning rate, batch size, number of epochs, and hardware; include these for reproducibility.","section":"Section III-A.3"},{"comment":"The clip-level branch is said to use 'two GMMFormer blocks,' while the implementation in Section III-A.3 lists four Gaussian variances (0.5, 1.0, 5.0, ∞); clarify the number of GMMFormer blocks and how the variances map to blocks.","section":"Section II-B"},{"comment":"The heatmaps lack labeled axes and a color scale; the caption states 'Darker regions indicate higher relevance' but the color mapping is not defined, making the visualization difficult to interpret.","section":"Figure 3"},{"comment":"The term 'Multi-Mamba' is not explained: Eq. (5) uses a single Mamba function, and the text says the module is integrated into both branches. Please define what makes it 'multi' (e.g., multiple Mamba blocks in parallel or in sequence) and describe the Mamba architecture (e.g., selective state space parameters) beyond the hyperparameters listed.","section":"Section II-D"}],"recommendation":"major_revision","confidential_remarks":"The paper presents an incremental combination of GMMFormer with a Mamba block and two cross-attention fusion modules. The central state-of-the-art claim is currently unverifiable because the label construction and training loss are unspecified, and the reported numbers do not consistently support the claim. The ablation results also contradict the narrative. These issues are fixable in a major revision, but the authors need to provide a rigorous experimental protocol and re-evaluate their claims. The paper would benefit from including the missing baselines and reporting variance; otherwise, the contribution is not clearly above the level of a workshop paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a modest, plausible extension of GMMFormer for PRVR — insert a Mamba block and two cross-attention fusion modules — with new benchmark numbers on three datasets. The writing is clear and the module descriptions are understandable. The ablation studies on Charades-STA mostly support the design, though not uniformly on R@1.\n\nWhat is genuinely new: the specific combination of Mamba with bidirectional temporal fusion on top of GMMFormer. That configuration is not in the prior work they cite. The paper also ships code, which is a point in its favor.\n\nThe soft spots are real and concentrated. The paper never states the training loss. Section II has no objective function. That alone makes the training setup unreproducible. More seriously, Section III-A.1 says the moment annotations 'were not used in our experiments.' In the standard PRVR protocol from MS-SL, those annotations define the relevance labels: a video is positive for a query only if the query matches one of its annotated moments. If they truly built no labels from the moment annotations, then there is no defined ground truth for ActivityNet Captions, Charades-STA, or TVR, and Table I is uninterpretable. The more charitable reading is that they meant the moment timestamps are not used at inference, while still deriving video-level positives from the annotations. But the text does not say that, and the exact pair-construction procedure is absent. This is a load-bearing omission, not a stylistic one.\n\nThe experimental claims are also weaker than the abstract suggests. On the primary R@1 metric, MamFusion is worse than GMMFormer on ActivityNet (8.0 vs 8.3) and Charades-STA (2.0 vs 2.1), and better only on TVR (14.2 vs 13.9). The paper leans on SumR, where it gains +1.4, +3.6, and +0.9, but it never declares SumR the primary criterion. The ablation table is internally inconsistent in places: removing V2T raises R@1 from 2.0 to 2.3, and removing both temporal fusions still gives R@1 2.1. The claim that every component is essential is overstated. Recent PRVR baselines DL-DKD and PEAN are cited in the introduction but do not appear in Table I. No variance or significance testing is reported.\n\nWho is this for: researchers actively working on PRVR or on Mamba-based video-text fusion. It is an incremental contribution, not a paradigm shift. With the missing training details supplied and the label-construction protocol clarified, the paper would be a reasonable workshop or modest conference paper. As written, the central SOTA claim is not verifiable.\n\nRecommendation: send to peer review, because the architecture is sensible and the code availability helps. But flag the missing loss and label protocol as the primary issues, and ask the authors to either release pair-generation code or state clearly how positives/negatives were built. A serious referee could turn this into a solid contribution.","headline":"Plausible incremental PRVR extension whose reported SOTA numbers rest on an unstated label-construction protocol; worth a careful review but needs major revisions.","tokens_in":10041,"tokens_out":2244,"would_cite":false,"duration_ms":20328,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"MamFusion claims state-of-the-art partially relevant video retrieval by combining a Multi-Mamba module with bidirectional temporal text-video fusion.","keywords":["partially relevant video retrieval","Mamba","state space model","temporal cross-modal fusion","video-text retrieval","long untrimmed video","GMMFormer","MamFusion"],"falsifier":"Re-run the released code with a fully specified training loss (for example, InfoNCE over video-query pairs) and an explicitly defined negative-sampling rule that does not peek at moment annotations, then compare MamFusion with GMMFormer under identical conditions; if the SumR gaps on ActivityNet Captions, Charades-STA, or TVR disappear or reverse, the central SOTA claim is settled negative.","tokens_in":9033,"feed_emoji":"🎬","tokens_out":6476,"duration_ms":68879,"temperature":0.7,"pith_summary":"This paper tries to show that partially relevant video retrieval—finding an untrimmed video in which only some moments match a text query—gets better when a long-sequence state-space model called Mamba is combined with two temporal cross-modal fusion modules. If the claim is right, MamFusion would be the top-ranked retrieval model in the paper's comparison on ActivityNet Captions, Charades-STA, and TVR, the standard benchmarks for this task. The practical motivation is that real video collections are long and mostly irrelevant to any given query, so retrieval models must identify the few relevant moments without expensive moment-by-moment search. The paper reports higher recall at every rank on Charades-STA and higher SumR on all three datasets than the GMMFormer baseline it extends.","feed_headline":"Mamba-based fusion claims new highs in partial video retrieval","feed_subtitle":"Selective state-space blocks plus bidirectional temporal fusion lift recall on three video benchmarks.","key_machinery":"The central object is the Multi-Mamba module: a selective state-space sequence model (Mamba) inserted after GMMFormer blocks in both the clip-level and video-level branches, which the paper says lets the network dynamically adjust to relevant moments while ignoring irrelevant frames. Around it sit two cross-attention fusion functions: Temporal V-to-T Fusion, where text tokens query the Mamba-processed video features, and Temporal T-to-V Fusion, where video features query the text representation. These are named in the paper; the machinery's job is to replace a purely global similarity score with temporal, query-aware alignment between text and video.","core_discovery":"The paper's central claim is that adding a Multi-Mamba module and two temporal fusion modules—Temporal V-to-T Fusion and Temporal T-to-V Fusion—on top of the GMMFormer clip and video branches produces a PRVR model that outperforms prior T2VR, VCMR, and PRVR methods. The Multi-Mamba module uses selective state-space processing to focus on query-relevant video segments and reduce redundancy in long untrimmed sequences; the two fusion modules use cross-attention in both directions so that the text representation is updated by video content and the video representation is updated by text content. On ActivityNet Captions, Charades-STA, and TVR, the reported SumR values are 147.4, 76.5, and 177.5, the highest in the comparison table. The paper further claims faster and smoother training loss convergence than GMMFormer, and ablation results indicate each module contributes to the final score.","pith_inferences":["The paper leaves the training loss and the construction of positive and negative video-query pairs unspecified, so a natural next step is to test whether the reported margins persist under an explicit contrastive loss with exactly specified negative sampling.","Because the moment annotations are deliberately unused, one testable extension is whether higher retrieval rank correlates with the annotated moment's presence at the top of the video—if the model finds videos with relevant moments but ranks them poorly in time, a localization-aware re-ranking could push recall further.","If replicated, the results suggest that long-video retrieval can move from quadratic transformer attention to linear-time state-space scanning, which would matter for hour-long real-world video collections."],"forward_implications":["If the reported numbers hold, MamFusion is the strongest published PRVR model on ActivityNet Captions, Charades-STA, and TVR by SumR.","Selective state-space models can serve as a drop-in replacement for the heavy clip-embedding machinery in PRVR, since the Multi-Mamba module compresses long sequences while preserving relevant moments.","The two temporal fusion directions provide a template for query-aware video scoring that other video-text retrieval systems could reuse without moment supervision.","The convergence-speed comparison implies the model reaches usable retrieval quality in fewer training epochs than the GMMFormer baseline, reducing training cost."],"supporting_citations":[{"why":"Defines the PRVR task and the R@K/SumR evaluation protocol this paper adopts.","marker":"[7]"},{"why":"Supplies the I3D visual features used for ActivityNet Captions.","marker":"[16]"},{"why":"GMMFormer is the baseline backbone that MamFusion builds on and the strongest PRVR comparator it reports beating.","marker":"[17]"},{"why":"Provides the bidirectional temporal fusion formulation that the two fusion modules adapt.","marker":"[20]"},{"why":"Supplies the pretrained language encoder used to extract sentence features.","marker":"[21]"},{"why":"Supplies the TVR dataset and its visual and language features.","marker":"[28]"},{"why":"Supplies the I3D features used for Charades-STA.","marker":"[29]"}],"fun_headline_variants":["Multi-Mamba with temporal fusion tops video retrieval","Mamba-based fusion sets new bar for partial video retrieval","MamFusion: state-space net excels at video retrieval","Temporal fusion and Mamba lift video retrieval recall"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper never states the training loss or the exact rule for turning moment-annotated videos into positive and negative video-query pairs, so the reported scores are only as trustworthy as that unstated protocol.","fun_headline_variants_meta":{"raw":{"variants":["Multi-Mamba with temporal fusion tops video retrieval","Mamba-based fusion sets new bar for partial video retrieval","MamFusion: state-space net excels at video retrieval","Temporal fusion and Mamba lift video retrieval recall"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000341,"raw_usage":{"total_tokens":1863,"prompt_tokens":916,"completion_tokens":947,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":532,"completion_tokens_details":{"reasoning_tokens":882}},"tokens_in":532,"tokens_out":947,"duration_ms":10200,"temperature":1.0,"reasoning_tokens":882,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:02:03.445742+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the released code with a fully specified training loss (for example, InfoNCE over video-query pairs) and an explicitly defined negative-sampling rule that does not peek at moment annotations, then compare MamFusion with GMMFormer under identical conditions; if the SumR gaps on ActivityNet Captions, Charades-STA, or TVR disappear or reverse, the central SOTA claim is settled negative.","supporting_citations":[{"cited_title":"Partially relevant video retrieval,","cited_arxiv_id":null,"evidence_quote":"Defines the PRVR task and the R@K/SumR evaluation protocol this paper adopts."},{"cited_title":"Gmmformer: Gaussian-mixture-model based transformer for efficient partially relevant video retrieval,","cited_arxiv_id":null,"evidence_quote":"GMMFormer is the baseline backbone that MamFusion builds on and the strongest PRVR comparator it reports beating."},{"cited_title":"Catr: Combinatorial-dependence audio-queried transformer for audio-visual video segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides the bidirectional temporal fusion formulation that the two fusion modules adapt."},{"cited_title":"Tvr: A large- scale dataset for video-subtitle moment retrieval,","cited_arxiv_id":null,"evidence_quote":"Supplies the TVR dataset and its visual and language features."},{"cited_title":"Local-global video-text interactions for temporal grounding,","cited_arxiv_id":null,"evidence_quote":"Supplies the I3D features used for Charades-STA."}],"review_version":1}