{"id":"5b00d97f-eee5-4fa1-8ac9-aebbf3fbebff","arxiv_id":"2506.07471","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"ARL detects ambiguous text-video pairs using uncertainty and similarity, then trains retrieval models with ambiguity-aware contrastive and triplet losses, achieving state-of-the-art on TVR and ActivityNet Captions.","lead":"Partially relevant video retrieval seeks videos that contain a segment matching a text query. This paper proposes detecting 'ambiguous' pairs that are neither clearly positive nor negative, and training the retrieval model to treat them flexibly, improving retrieval accuracy on standard benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Validation gap in ambiguity detection: unverified that Eq. 6–7 finds genuine relevance rather than hard negatives.","rationale":"I agree with the reader's weakest assumption and would keep the verdict unchanged. The paper is a substantial engineering contribution: component ablations are consistent, the SumR improvement over GMMFormer is large on TVR (185.9 vs 176.6 in Table 1), and the model complexity is competitive. The unresolved point is causal: whether the improvement comes from modeling genuine label ambiguity or from a generic reweighting of hard negatives. That can be settled only by checking the detector's output against an external relevance signal. I would keep the reader's conditional verdict, with code release, multi-seed variance, and the validation protocol above as prerequisites. One secondary correction: 'outperforms previous works in all recall metrics' is not literally true—on ActivityNet R@5 in Table 1, Ours is 24.6 vs GMMFormer's 24.9—but this does not change the SumR-based SOTA claim. The self-referential detection loop and the threshold selection (mean positive similarity / mean uncertainty) are exactly why external validation, not just component ablation, is necessary.","tokens_in":12364,"tokens_out":13183,"duration_ms":181199,"concrete_test":"Run an external precision study: after training ARL, sample 200 unpaired text-video pairs flagged as ambiguous by Eq. 6–7 (using the retrieval branch) and 200 unpaired pairs below threshold. Have at least three annotators judge, under the PRVR definition, whether the video contains a segment matching the query; majority vote is the gold label. Report precision, recall, AUC of the ambiguity score against the gold labels, and inter-annotator agreement. If AUC is near 0.5 or precision is not substantially above the unpaired base rate, the detector is not finding genuine ambiguity and the SumR gains should be attributed to hard-negative relaxation rather than to the proposed mechanism. On TVR, a cheaper proxy is available: evaluate T-F LAD's detected frames inside paired videos against the annotated moment spans.","verdict_should_be":"UNCHANGED","load_bearing_attack":"ARL's central mechanism is the LAD detector in Eqs. 6–7: a text-video pair is called ambiguous when model-estimated similarity s and uncertainty u both exceed data-dependent thresholds, and that detected set is then used as soft supervision in Eqs. 8–12. For the claim that ARL 'addresses label ambiguity' to hold, high s and high u must pick out unpaired videos that genuinely contain a segment matching the query, not merely hard negatives. The paper's evidence for this is Fig. 4 (distributions), Fig. 5 (two qualitative examples), and Table 6 (an ablation over how to consume the detected set); none measures whether the detected pairs are actually relevant. Because detection and training use the same online embeddings, the loop can improve SumR simply by relaxing high-similarity negatives, with no semantic ambiguity involved. That would leave the reported numbers intact but invalidate the proposed explanation. This validation gap is the weakest load-bearing assumption in the paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Ambiguity-Restrained Representation Learning (ARL) for Partially Relevant Video Retrieval (PRVR). ARL detects ambiguous text-video pairs using two criteria: uncertainty, defined as the average similarity of an instance to all instances in the dataset, and pairwise similarity, defined as the maximum text-frame similarity. Detected ambiguous pairs are then treated neither as hard negatives nor as positives; instead they are incorporated into a multi-positive contrastive loss and a dual triplet margin loss at both text-video and text-frame levels. A cross-model mechanism exchanges detected ambiguous sets between two identical encoders to reduce error propagation. Experiments on TVR and ActivityNet Captions report improved SumR over prior PRVR methods, with ablations showing contributions from each component.","tokens_in":12576,"tokens_out":3785,"duration_ms":45720,"significance":"If the ambiguity-detection mechanism is valid, the paper offers a new and potentially useful perspective on PRVR: pair-wise labeling in retrieval datasets creates ambiguous relations that should be modeled softly rather than as negatives. The reported TVR improvements are meaningful, and the ablation study in Table 5 supports the contribution of each component. The paper also provides distribution analyses and qualitative examples. However, the central mechanism is not directly validated: the paper does not measure whether the detected ambiguous pairs are actually relevant to the query, and the ActivityNet gains are small and inconsistent across recall metrics. The final retrieval evaluation is external, so the reported numbers are not circular, but the proposed explanation for the gains is underdetermined. With additional validation experiments, this could be a solid contribution; in its current form, the main claim rests on an unverified assumption.","major_comments":[{"comment":"The central claim that the detected ambiguous sets are genuinely relevant, rather than hard negatives, is not validated. Equations (6)–(7) define ambiguous pairs using the model's own similarity and uncertainty estimates, and the evidence for their semantic validity is limited to Fig. 4 (distributions), Fig. 5 (two qualitative examples), and Table 6 (an ablation over how to consume the detected set). None of these measures whether the detected unpaired videos contain a segment that matches the query. Because detection and training use the same online embeddings, the method could improve SumR simply by relaxing high-similarity negatives, without any semantic ambiguity being involved. A concrete fix is to use the moment/segment annotations available in TVR and ActivityNet Captions: for each detected ambiguous pair, check whether the video contains a segment matching the query, and report precision/recall of detection over training epochs. This would directly test the load-bearing assumption.","section":"Method, Eqs. (6)–(7); Experiments, Qualitative Results"},{"comment":"The statement that the method 'outperforms previous works in all recall metrics' is not supported by Table 1. On ActivityNet Captions with ResNet/I3D/RoBERTa features, the proposed method achieves R@5 = 24.6, which is lower than GMMFormer's 24.9, and R@1 is tied at 8.3. The SumR improvement is 2.3, but no error bars, multiple seeds, or significance tests are reported. Given that the ActivityNet gains are marginal and inconsistent across metrics, the claim of state-of-the-art performance on both datasets requires repeated runs and statistical testing.","section":"Tables 1 and 2"},{"comment":"The thresholds τ_s and τ_u that define ambiguous pairs are set per epoch using statistics of the training set (mean of the positive-pair similarity distribution and mean of the uncertainty distribution). The paper provides no sensitivity analysis for these thresholds. Since the entire method depends on the detected ambiguous sets, the results could be sensitive to the choice of threshold. The authors should report performance as a function of τ_s and τ_u, or use a validation-based selection procedure, to demonstrate that the reported gains are not brittle.","section":"Implementation Details"},{"comment":"The claim that reciprocal pseudo-label exchange between two identical encoders 'mitigates error propagation' is not directly tested. Table 5 rows (d) and (e) only compare the full model with and without cross-model detection, which shows a SumR improvement but does not isolate error-propagation mitigation from ensemble or regularization effects. A more direct test would compare single-model self-training with the same architecture against the proposed cross-model variant, and track the precision of detected ambiguous pairs over epochs to see whether error accumulation is actually reduced.","section":"Cross-Model Ambiguity Detection"}],"minor_comments":[{"comment":"The phrase 'we achieves state-of-the-art performance' should be corrected to 'we achieve state-of-the-art performance'.","section":"Abstract and Contributions"},{"comment":"The word 'repec- tively' is a typo and should be 'respectively'.","section":"Eq. (3)"},{"comment":"The set notation in the denominators, e.g., 'V ∈ A^q_i ∨ N^q_i', is ambiguous; it should be written as a union of sets, e.g., 'V ∈ A^q_i ∪ N^q_i', with explicit definitions of how negatives are sampled.","section":"Eqs. (8)–(10)"},{"comment":"The figure caption and diagram use several arrows and 'Logical AND' symbols without explaining the data flow clearly; please label the components or refer to specific equations in the caption.","section":"Fig. 3"},{"comment":"The column header 'A' is undefined; it should indicate the way ambiguous sets are used (e.g., 'Positive', 'Ignore', 'Ours').","section":"Table 6"},{"comment":"The text states 'More details are provided in the appendix,' but the appendix is not included in the manuscript; please include it or remove the reference.","section":"Appendices"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the validation gap for the ambiguity detector. If the authors can provide annotation-based evaluation of detected ambiguous pairs (e.g., using TVR and ActivityNet segment annotations) and add significance testing for the ActivityNet results, the paper would be substantially stronger. The current evidence is suggestive but not conclusive."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead the PRVR paper from Cho et al. (arXiv:2506.07471). The core claim is that pairwise labeling in untrimmed video retrieval creates ambiguous text-video pairs, and the paper handles them by detecting ambiguity from the model's own similarity and uncertainty estimates, then softening the loss for those pairs. The idea is genuinely new in the PRVR literature; prior work treats all unpaired pairs as hard negatives. The paper builds this out with multi-positive contrastive, dual triplet margins, and a text-frame level extension, and reports solid gains on TVR (9.3 SumR over GMMFormer with ResNet/I3D/RoBERTa) and smaller gains on ActivityNet. The ablations in Tables 5-6 are consistent with the design; the cross-model detection is a reasonable attempt to break the self-training loop.\n\nThe main soft spot is exactly what the stress-test flagged: there is no direct verification that the detected ambiguous pairs are actually semantically relevant to the query rather than just high-similarity hard negatives. Fig. 4 shows the distributions separate but not that the high-similarity/high-uncertainty set contains actual positives. Fig. 5 shows two cherry-picked examples. Table 6 compares ways to consume the ambiguous set but never measures its precision. Since the detector and the trained model share the same embeddings, the SumR gains could partly come from simply relaxing high-similarity negatives. The cross-model exchange mitigates but doesn't eliminate this. Also missing: error bars or multi-seed runs, and the thresholds are dataset-fitted (mean of positive similarity, mean of uncertainty), so the 9.3% could be partially threshold-tuning artifact. On ActivityNet the margin is small.\n\nThat said, the central argument is plausible and the engineering is clean. The paper is a legitimate contribution to PRVR, not a breakthrough but a solid step. It deserves a serious referee. I'd recommend acceptance if the authors can add variance numbers and at least a small-scale human or external evaluation of the detected ambiguous set. Without that, the SOTA claim is not fully established but the method is still useful.\n\nWho is this for: people working on video retrieval, noisy labels in cross-modal learning, or self-training. I'd bring it to reading group.\n\nRecommendation: engage with it; ask for the missing validation before fully buying the claim.","headline":"A plausible, well-engineered method for treating ambiguous text-video pairs in PRVR, but the central ambiguity-detection mechanism lacks direct validation and the SOTA claim needs variance and external checks.","tokens_in":13066,"tokens_out":1649,"would_cite":true,"duration_ms":16204,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Partially relevant video retrieval fails when every unpaired video is treated as a hard negative; this paper detects ambiguous pairs and relaxes their training constraint, reporting state-of-the-art recall on TVR and ActivityNet.","keywords":["partially relevant video retrieval","text-video retrieval","label ambiguity","uncertainty estimation","contrastive learning","multi-positive learning","dual triplet margin loss","cross-model learning"],"falsifier":"Annotate a held-out set of unpaired text-video pairs from TVR or ActivityNet as truly relevant or truly irrelevant; if the pairs ARL flags as ambiguous are mostly judged irrelevant by human raters, or if removing the ambiguity relaxation (training them as hard negatives) does not lower recall, the central claim is falsified.","tokens_in":12186,"feed_emoji":"🎬","tokens_out":6539,"duration_ms":70361,"temperature":0.7,"pith_summary":"Partially relevant video retrieval asks a model to find untrimmed videos in which some segment matches a text query, yet standard training labels every unpaired video as a hard negative. This paper argues that such pair-wise labels are often wrong in spirit: an unpaired video can share the query's context even when it is not the annotated match, so forcing it away from the query corrupts the learned representation. To counter this, ARL first flags ambiguous text-video pairs using two complementary signals—dataset-level uncertainty (how much a query or video overlaps with many other instances) and pair-level similarity (the best matching frame)—then trains with those pairs treated neither as full positives nor as hard negatives, via multi-positive contrastive learning and a dual triplet margin loss. The method extends the same relaxation down to individual frames inside a video and uses two symmetric encoders that exchange their detected ambiguous sets to keep errors from compounding. On TVR and ActivityNet Captions, the reported result is a state-of-the-art retrieval score, with SumR (sum of recall at ranks 1, 5, 10, and 100) gains of 9.3 and 2.3 points over the prior best model in the ResNet/I3D/RoBERTa setup.","feed_headline":"Soft-labeling ambiguous pairs lifts video retrieval by 9.3%","feed_subtitle":"Videos that share a query's context become soft positives rather than negatives, setting new bests on TVR and ActivityNet.","key_machinery":"The mechanism that carries the argument is a pair of Label Ambiguity Detection (LAD) modules operating at two granularities. Text-video LAD computes a dataset-wise uncertainty for every query and video (the average similarity of that instance to all instances of the other modality) and a pair-level similarity for each query-video pair (the maximum cosine similarity between the query and the video's frames), then flags a pair as ambiguous when both values exceed thresholds set to the mean of the training distributions. Text-frame LAD applies the same test between a query and the frames of a single untrimmed video. Detected ambiguous sets are fed into a multi-positive supervised contrastive loss and a dual triplet margin loss with a smaller margin for ambiguous samples than for true negatives, and a two-encoder cross-model scheme lets each encoder train on the other's detected ambiguous sets.","core_discovery":"The paper's central claim is that the one-to-one labeling convention used to train PRVR models—one positive video per query, every other video negative—actively harms retrieval because untrimmed videos routinely share context with unpaired queries. It asserts that this ambiguity can be detected online with two complementary metrics, uncertainty and similarity, and that restraining the loss on detected ambiguous pairs instead of treating them as negatives yields better text-video alignment. On the paper's own terms, ARL is the first method to make this ambiguity explicit in PRVR, and it reports state-of-the-art recall on both TVR and ActivityNet Captions, including a 9.3-point SumR improvement over the strongest prior model in the standard feature setting and 9.1- and 3.3-point gains with CLIP-L/14 features.","pith_inferences":["This suggests the same two-signal ambiguity test could be used to audit other pairwise-labeled retrieval datasets: instances with high dataset-level uncertainty and high pair similarity are candidate false negatives, and re-weighting them as soft positives might improve retrieval there too.","A natural stress test is to replace the full-dataset uncertainty computation with an online or approximate estimate; if results hold, the method scales to much larger corpora where an exact similarity map over all pairs each epoch is costly.","The text-frame ambiguity formulation points toward weakly supervised video moment grounding: the frames that survive the ambiguity test inside a positive video are, in effect, pseudo-labels for the relevant segment."],"forward_implications":["On TVR and ActivityNet Captions, ARL achieves higher R@1, R@5, R@10, R@100, and SumR than prior PRVR methods under the same features, with the largest reported margin being a 9.3-point SumR gain in the ResNet/I3D/RoBERTa setting.","The text-frame ambiguity module lets the model use information from every frame of an untrimmed video, not just the single best-matching frame, without any additional annotation.","Cross-model ambiguity detection prevents one model from repeatedly reinforcing its own mistaken ambiguity predictions, so the training signal stays cleaner as training progresses.","Because the thresholds are set each epoch from the mean of the training similarity and uncertainty distributions, the method adapts to the model's changing representation rather than relying on fixed cutoffs."],"supporting_citations":[{"why":"Establishes the PRVR task and the MS-SL baseline that ARL must beat, and supplies the multi-scale clip modeling approach that ARL critiques.","marker":"(Dong et al. 2022)"},{"why":"Introduces GMMFormer, the strongest prior PRVR model on which the central comparison and efficiency trade-offs are made.","marker":"(Wang et al. 2024)"},{"why":"DL-DKD is the knowledge-distillation baseline that ARL outperforms even without an additional teacher model.","marker":"(Dong et al. 2023)"},{"why":"Provides the TVR dataset and the video corpus moment retrieval setting used for evaluation.","marker":"(Lei et al. 2020)"},{"why":"Provides the ActivityNet Captions dataset used as the second benchmark.","marker":"(Krishna et al. 2017)"},{"why":"Supplies the CLIP-L/14 pretrained features used in the large-model experiments.","marker":"(Radford et al. 2021)"},{"why":"Motivates cross-model ambiguity detection through the co-teaching idea of two networks training each other on noisy labels.","marker":"(Han et al. 2018)"},{"why":"Supplies the supervised contrastive objective that ARL modifies into its multi-positive form.","marker":"(Khosla et al. 2020)"}],"fun_headline_variants":["Ambiguous pairs as soft positives boost video retrieval","Rethinking negatives in text-video retrieval improves recall","ARL: detect and restrain ambiguity for better video search","Soft-labeling ambiguous pairs lifts retrieval performance","New training strategy for partially relevant video retrieval"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method depends on the assumption that a pair it flags as ambiguous—high average overlap plus high pairwise similarity—really is a relevant-but-unlabeled pair, and that the chosen mean thresholds separate those from ordinary negatives.","fun_headline_variants_meta":{"raw":{"variants":["Ambiguous pairs as soft positives boost video retrieval","Rethinking negatives in text-video retrieval improves recall","ARL: detect and restrain ambiguity for better video search","Soft-labeling ambiguous pairs lifts retrieval performance","New training strategy for partially relevant video retrieval"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000175,"raw_usage":{"total_tokens":1295,"prompt_tokens":967,"completion_tokens":328,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":583,"completion_tokens_details":{"reasoning_tokens":256}},"tokens_in":583,"tokens_out":328,"duration_ms":4472,"temperature":1.0,"reasoning_tokens":256,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:32:33.121914+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Annotate a held-out set of unpaired text-video pairs from TVR or ActivityNet as truly relevant or truly irrelevant; if the pairs ARL flags as ambiguous are mostly judged irrelevant by human raters, or if removing the ambiguity relaxation (training them as hard negatives) does not lower recall, the central claim is falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the PRVR task and the MS-SL baseline that ARL must beat, and supplies the multi-scale clip modeling approach that ARL critiques."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces GMMFormer, the strongest prior PRVR model on which the central comparison and efficiency trade-offs are made."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DL-DKD is the knowledge-distillation baseline that ARL outperforms even without an additional teacher model."},{"cited_title":"L.; and Bansal, M","cited_arxiv_id":null,"evidence_quote":"Provides the TVR dataset and the video corpus moment retrieval setting used for evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Motivates cross-model ambiguity detection through the co-teaching idea of two networks training each other on noisy labels."}],"review_version":1}