{"id":"5bbb224c-4ccb-4dbd-9566-c1ebb426d1f9","arxiv_id":"2505.06557","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A positive sample mining method that groups training queries by text similarity and uses contrastive plus rank losses improves weakly supervised temporal sentence grounding on Charades-STA, ActivityNet Captions, and NExT-GQA.","lead":"This paper adds a training-time module to weakly supervised video moment grounding that finds query-video pairs with similar text descriptions and treats them as positive samples instead of negatives. On three benchmarks it improves grounding accuracy by a few points over strong baselines, with minimal extra inference cost.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central mechanism assumes top-k text-similar queries have semantically similar target moments; the paper never verifies this, and its own limitations admit it often fails.","rationale":"The reader's conditional verdict already centers on the same assumption: text-query cosine similarity is treated as a reliable proxy for proposal/moment similarity. My stress-test agrees that this is the most load-bearing point, because both proposed losses are defined relative to the mined similar/dissimilar split; if that split is noisy, the method's stated mechanism is undermined even though the aggregate numbers may still improve as a regularizer. The rank-loss proxy in Eqs. (7)-(12) is a second, real weakness: L_query and L_prop are margin violations with respect to a randomly chosen dissimilar sample, not distances, so 'lower contrastive loss' is only weakly related to closeness; however, this is secondary because an unconditional version of the contrastive loss would carry much of the method even if the rank loss were removed. The concrete random-positive control would distinguish semantic mining from generic inter-sample alignment, and the CLIP-segment similarity check would directly measure the text-to-moment assumption. I do not see grounds to move from CONDITIONAL to ACCEPT or REJECT: the reported gains are internally consistent and ablations support the components, but the missing verification of the core assumption and the absent code/error bars justify keeping the verdict conditional. Therefore the reader's verdict stands unchanged.","tokens_in":22440,"tokens_out":6140,"duration_ms":64108,"concrete_test":"Run a control on Charades-STA: keep every component of PPS+PSM identical, but replace the top-k similar subset in Eqs. (1)-(3) with k randomly selected training samples. If R@1, IoU=0.5 and mIoU stay close to 54.36/46.92, the gain is not attributable to semantic positive mining and the central claim fails; if they drop back toward the 51.49/44.87 baseline, semantic selection matters. Additionally, on a small labeled subset, compute the CLIP cosine similarity between the anchor's ground-truth segment and each top-k similar sample's ground-truth segment and compare it with the same similarity for random pairs; a null result would confirm that text similarity does not proxy moment similarity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"PSM's two losses both depend on the mined 'similar' subset from Eqs. (1)-(3), which is selected purely by SentenceTransformer cosine similarity of the text queries. The paper's premise is that text-similar queries imply semantically similar proposals/moments, so pulling p toward p_sim/q_sim and ranking p above p_n with respect to q_sim/p_sim is correct supervision. No experiment in the paper tests this premise. In Charades-STA, short queries with shared vocabulary can denote different events ('open' vs 'close' the refrigerator; 'walk to' vs 'sit at' a table); in ActivityNet/NExT-GQA, long complex queries often have no reliable similar samples, as Sec. V and Fig. 6(d) concede. If a large share of top-k neighbors are false positives, the PSM-guided contrastive and rank losses actively pull the anchor proposal toward unrelated temporal intervals, so the reported gains may come from generic inter-sample regularization rather than from the claimed positive-sample semantics. The only direct evidence is the t-SNE figure with two hand-picked anchors, which cannot quantify the false-positive rate. This is a correctness risk in the causal claim, not a disagreement with external consensus: the method may work, but for a different reason than stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Positive Sample Mining (PSM) for weakly supervised temporal sentence grounding (WSTSG). For each anchor query-video sample, PSM partitions the remaining training samples into similar and dissimilar subsets using SentenceTransformer cosine similarity of the text queries (Eqs. (1)-(3)). It then adds a PSM-guided contrastive loss, which pulls the anchor proposal toward the mined similar sample and pushes it away from the dissimilar sample, and a PSM-guided rank loss, which encourages the anchor proposal to be closer to the similar sample than the intra-video negative proposal. The method is evaluated on Charades-STA, ActivityNet Captions, and NExT-GQA, showing consistent improvements over PPS and NG baselines, with ablations on each loss term, the number of similar samples, feature types, backbones, and hyperparameters. The paper claims this is the first positive-sample-mining approach for WSTSG and that it also improves grounded VideoQA without requiring additional annotations.","tokens_in":22656,"tokens_out":8647,"duration_ms":86647,"significance":"If the mechanism holds, PSM provides a new source of discriminative supervision for WSTSG by exploiting cross-sample semantic correlations instead of treating all non-anchor samples as negatives. The paper is thorough on the experimental side: it validates on three benchmarks, integrates PSM with CNM, CPL, and PPS, ablates each component, and shows that the mining overhead is small relative to training cost. The NExT-GQA results are interesting because PSM improves grounding and QA over NG+ without the GPT-4 annotations used by NG+. The central claim is falsifiable, and the main open questions concern whether the mined samples are truly positive and whether the rank loss implements the stated distance objective; these are empirically and analytically checkable. If the authors resolve those questions, the contribution would be a useful addition to weakly supervised video-language grounding.","major_comments":[{"comment":"The key premise of the method is that samples whose queries are textually similar to the anchor query have target moments that are semantically similar to the anchor moment. This premise is introduced in Eqs. (1)-(3), where the similar subset is selected purely by SentenceTransformer cosine similarity of text queries, but it is never measured or validated. The only direct evidence is the two hand-picked examples in Fig. 5, and the paper's own Limitations section (Sec. V) states that complex queries often have no reliable similar samples, with Fig. 6(d) showing a failure case. Because both PSM losses are applied only to the mined subset, a non-negligible false-positive rate would make the losses actively pull the anchor proposal toward unrelated temporal intervals, and the gains in Tables I-III could stem from generic inter-sample regularization rather than from positive-sample semantics. Please provide a quantitative validation of the mining: for a random sample of anchors on Charades-STA and ActivityNet Captions, compute the tIoU between the ground-truth moments of the anchor and the ground-truth moments of the top-k mined samples (or have a subset human-annotated), report precision@k for the k values considered in Table VI, and compare against a control with randomly selected 'similar' sets of the same size. This would directly test the causal claim in Sec. IV-D that the improvement is primarily due to discriminative supervision from positive samples.","section":"III-B, Eqs. (1)-(3); Sec. V"},{"comment":"In Eqs. (7)-(12), the paper sets d(p,q_sim)=L_query and d(p,p_sim)=L_prop, i.e., it uses the clipped margin losses as distances. However, L_query = max(p·q_dis - p·q_sim + γ1, 0) measures how much the anchor proposal fails to be more similar to q_sim than to q_dis by margin γ1; it is not a distance to q_sim. The same holds for L_prop. Consequently, the rank loss in Eq. (11) does not implement the stated objective that similar samples are closer to the anchor proposal than to the negative intra-video proposal: when both L_query and L^n_query are zero, L^query_rank equals the constant γ5 with zero gradient; when both are nonzero, the loss can be reduced by lowering p·q_dis without increasing p·q_sim. The supporting statement in Sec. III-D that 'a lower PSM-guided contrastive loss indicates that the intra-video proposal is closer to similar samples' is therefore not supported. I recommend either reformulating the rank loss with an actual similarity/distance measure (e.g., max(p·q_sim - pn·q_sim + γ, 0)) and reporting whether the gains in Table IV persist, or providing a rigorous analysis of the hinge-based distance, including what happens in the inactive and double-active regions. The CR-loss ablation in Table V is related but does not by itself justify the hinge-based quantity as a distance.","section":"III-D, Eqs. (7)-(12)"}],"minor_comments":[{"comment":"Equation (12) writes L^prop_rank with d(p,q_sim)-d(pn,q_sim) but then substitutes L_prop - L^n_prop; based on Eq. (10), the first arguments should be p_sim, not q_sim. Please fix the typo.","section":"III-D, Eq. (12)"},{"comment":"In the ActivityNet results paragraph, five metrics are listed (R@1 IoU=0.5, R@1 IoU=0.7, R@5 IoU=0.3, R@5 IoU=0.5, R@5 IoU=0.7) but only four percentage gains are reported; please correct the enumeration.","section":"IV-D"},{"comment":"The caption mentions 'PSM-guided contrastive loss' twice in the final sentence; the second occurrence appears to mean 'PSM-guided rank loss' and should be corrected.","section":"Fig. 2 caption"},{"comment":"The main results are reported without variance across runs; given that several improvements are in the 1-3 point range, please report multiple-seed means and standard deviations, or state explicitly if this is not standard practice for the baselines used.","section":"Tables I-III"},{"comment":"The row 'Ours (NG + PSM)' is based on the NG baseline, while the comparison row 'NG+' uses additional GPT-4 annotations; please make this explicit in the table caption or in a footnote so that the comparison is not read as using the same training signal.","section":"Table III"}],"recommendation":"major_revision","confidential_remarks":"The paper falls within the scope of TCSVT and the experimental work is substantial. I believe the empirical gains are credible, but the two major comments concern the paper's stated mechanism rather than mere presentation: the mining premise needs a direct validation experiment, and the rank loss needs either reformulation or a rigorous justification of the hinge-as-distance choice. Both are addressable within the scope of a revision, so I would support acceptance if these points are adequately resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper deserves a proper review. It reports a simple, training-time positive-sample mining module for weakly supervised temporal sentence grounding that gives consistent gains across Charades-STA, ActivityNet, and NExT-GQA, and across three backbones. The headline numbers are credible: PPS+PSM goes from 51.49 to 54.36 on Charades-STA R@1@0.5, and the ablation tables show each loss component contributing. I believe the empirical result.\n\nWhat's genuinely new: this is the first application of inter-sample positive mining to WSTSG. The authors partition the training set by SentenceTransformer query similarity, then use a margin contrastive loss to pull the anchor proposal toward similar samples and push it from dissimilar ones, plus a rank loss that makes similar samples closer to the anchor proposal than to an intra-video negative. The idea is clean, and the implementation is cheap.\n\nThe soft spots are real but not fatal. The biggest is that the paper never verifies its load-bearing premise: that text-similar queries actually refer to similar temporal moments. The shared-vocabulary counterexample ('open the refrigerator' vs 'close the refrigerator') is obvious, and the authors themselves concede in Sec. V that complex queries often lack reliable similar samples. The t-SNE figure is just two hand-picked anchors. So the reported gains may come from generic inter-sample regularization rather than from semantically meaningful positive pulls. The ablations against hard-negative mining and a cosine-similarity rank loss partly address this, but they don't directly measure the false-positive rate of the mined 'positives.'\n\nTwo smaller technical issues. The rank loss treats the hinge losses in Eqs. (7)-(10) as distance metrics, which is an odd choice—once the contrastive margin is satisfied, the 'distance' is clipped to zero and the rank loss's behavior changes. Eq. (12) also has a typo: the first expression refers to distances to p_sim, the second to L_prop. Both should be fixed. And there are no error bars or variance seeds, which is common in this area but worth asking for.\n\nOverall: this is a useful, honest paper. It should go to peer review. I'd suggest the reviewers push for either a quantitative sanity check on the mined positives (e.g., how often do top-k text neighbors have IoU-overlapping ground-truth moments?) or a reframing of the method as a regularizer rather than as semantic positive mining.","headline":"A solid, honest WSTSG plug-in with consistent gains, but the 'positive sample' premise is asserted rather than demonstrated.","tokens_in":23231,"tokens_out":2893,"would_cite":true,"duration_ms":25681,"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":"In weakly supervised temporal grounding, mining text-similar training clips as positives—rather than pushing them away—produces more discriminative supervision and higher localization accuracy.","keywords":["weakly supervised temporal sentence grounding","positive sample mining","contrastive learning","rank loss","video-language grounding","grounded video question answering","text embedding similarity"],"falsifier":"Select a subset of a grounding dataset where each anchor's top-k text-similar queries are intentionally chosen to describe different moments from the anchor (same objects and verbs but different sub-events), train with PSM, and compare grounding accuracy on that subset against the no-PSM baseline; a drop would show the similarity assumption is violated. A simpler observation: compute the ground-truth temporal IoU between anchor moments and their mined similar samples; if a large share have near-zero IoU, the pull-close term is being supervised toward wrong moments.","tokens_in":22229,"feed_emoji":"🎯","tokens_out":9256,"duration_ms":82208,"temperature":0.7,"pith_summary":"This paper claims that weakly supervised temporal sentence grounding—locating the interval a sentence describes in an untrimmed video using only video-level text-video pairs—improves when the training set is not treated as uniformly negative. Its Positive Sample Mining (PSM) method partitions the other training samples into text-similar and text-dissimilar subsets, then uses two margin-based losses to pull the anchor proposal toward similar samples and away from dissimilar ones. On Charades-STA, adding PSM to the PPS baseline raises R@1 at IoU=0.5 from 51.49 to 54.36 and at IoU=0.7 from 26.16 to 28.73, with gains also reported on ActivityNet Captions and on the NExT-GQA grounded video question answering benchmark. The practical point is that semantically related clips carry alignment signal that existing contrastive schemes throw away.","feed_headline":"Training with similar clips as positives lifts grounding by 2.9 points","feed_subtitle":"Pulling text-similar clips close, not away, helps models locate moments without timestamp labels.","key_machinery":"The load-bearing object is the similar/dissimilar partition of the training set, computed by cosine similarity of SentenceTransformer text embeddings. For an anchor query $Q_i$, the top-$k$ most similar queries form the positive set $S_i^{\\mathrm{sim}}$ and all remaining samples form $S_i^{\\mathrm{dis}}$. Two margin losses then carry the signal: the PSM-guided contrastive loss $\\mathcal{L}^{\\mathrm{CL}}_{\\mathrm{PSM}}$ enforces $p\\cdot q_{\\mathrm{dis}} - p\\cdot q_{\\mathrm{sim}} + \\gamma_1 \\le 0$ (and the analogous proposal term), and the PSM-guided rank loss $\\mathcal{L}^{\\mathrm{rank}}_{\\mathrm{PSM}}$ reuses those contrastive values as distance measures so that the similar sample is closer to the anchor proposal than to the negative intra-video proposal. The machinery matters because it converts a text-similarity structure into local proposal-space supervision without any timestamp labels.","core_discovery":"The central claim is that some remaining training samples are positive evidence, not noise: for each anchor query-video pair, PSM computes cosine similarity between the anchor text query and all other queries using pretrained SentenceTransformer features, keeps the top-k as a similar subset and the rest as a dissimilar subset, and then supervises the proposal encoder with two losses. The PSM-guided contrastive loss makes the anchor proposal closer to a similar sample than to a dissimilar sample, in both query and proposal modalities. The PSM-guided rank loss makes a similar sample closer to the anchor proposal than to the negative intra-video proposal, which sharpens the contrast between the predicted moment and confusing segments within the same video. The paper reports consistent improvements across the CNM, CPL, and PPS backbones, and its ablations show the two losses are complementary.","pith_inferences":["Inference: the same recipe could transfer to other weakly supervised video-language tasks, such as text-to-video retrieval or weakly supervised temporal action localization, wherever contrastive losses currently repel all cross-sample pairs.","Inference: because PSM outperforms a method that relies on GPT-4-generated annotations, cross-sample mining may serve as a general substitute for LLM pseudo-labels in grounded QA and similar annotation-hungry settings.","Inference: the fixed top-$k$ cutoff leaves room for an adaptive or confidence-weighted similar set; the paper's own ablation shows performance saturates around $k=20$, hinting that noisy members begin to contaminate the positive set.","Inference: a natural extension is sub-event-level mining for complex queries, which the paper names as its main limitation; retrieving positives for decomposed query fragments could directly attack the case where whole-sentence neighbors are unreliable."],"forward_implications":["On Charades-STA, PPS+PSM raises R@1 at IoU=0.5 from 51.49 to 54.36 and R@1 at IoU=0.7 from 26.16 to 28.73.","On ActivityNet Captions, PPS+PSM raises R@1 at IoU=0.5 from 59.29 to 60.81 and improves most R@5 metrics across IoU thresholds.","On NExT-GQA, NG+PSM improves mIoP, mIoU, and Acc@GQA over both the NG baseline and the NG+ variant, without needing GPT-4 annotations.","Ablations on CNM, CPL, and PPS show the scheme transfers across different backbone proposal generators, and the training-time-only design adds no inference cost.","The two losses are complementary: the ablation table's best configuration uses all four loss terms together."],"supporting_citations":[{"why":"PPS is the base WSTSG model that PSM is built on; its Gaussian-mixture proposals and pull-push learning provide the anchor and negative intra-video proposals.","marker":"[17]"},{"why":"CPL supplies the transformer-based proposal and query encoder architecture and the multi-anchor proposal selection scheme that PSM inherits.","marker":"[16]"},{"why":"CNM is the contrastive negative-sample-mining baseline and one of the backbones used to show PSM generalizes.","marker":"[23]"},{"why":"SentenceTransformer supplies the text embeddings whose cosine similarity defines the similar and dissimilar sample partition in Eqs. (1)-(3).","marker":"[66]"},{"why":"NG and NG+ are the grounded VideoQA baselines on NExT-GQA that PSM is built on and compared against.","marker":"[68]"},{"why":"Charades-STA provides the primary benchmark and the R@n, IoU=m evaluation protocol used in the main experiments and ablations.","marker":"[10]"}],"fun_headline_variants":["Mining similar clips as positives boosts temporal grounding","Positive sample mining sharpens video moment localization","Using similar clips as positives, not negatives, lifts grounding","Rewriting negatives as positives improves weakly supervised grounding"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that if two text queries are close in SentenceTransformer embedding space, their corresponding video moments should also be close in proposal space; that equivalence can fail when sentences share words or style but describe different events, and the paper concedes that complex queries often have no reliable similar samples.","fun_headline_variants_meta":{"raw":{"variants":["Mining similar clips as positives boosts temporal grounding","Positive sample mining sharpens video moment localization","Using similar clips as positives, not negatives, lifts grounding","Rewriting negatives as positives improves weakly supervised grounding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00071,"raw_usage":{"total_tokens":3193,"prompt_tokens":935,"completion_tokens":2258,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":2198}},"tokens_in":551,"tokens_out":2258,"duration_ms":13885,"temperature":1.0,"reasoning_tokens":2198,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:38:28.913988+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Select a subset of a grounding dataset where each anchor's top-k text-similar queries are intentionally chosen to describe different moments from the anchor (same objects and verbs but different sub-events), train with PSM, and compare grounding accuracy on that subset against the no-PSM baseline; a drop would show the similarity assumption is violated. A simpler observation: compute the ground-truth temporal IoU between anchor moments and their mined similar samples; if a large share have near-zero IoU, the pull-close term is being supervised toward wrong moments.","supporting_citations":[{"cited_title":"Weakly supervised video moment localization with contrastive negative sample mining","cited_arxiv_id":null,"evidence_quote":"CNM is the contrastive negative-sample-mining baseline and one of the backbones used to show PSM generalizes."},{"cited_title":"Can i trust your answer? visually grounded video question answering","cited_arxiv_id":null,"evidence_quote":"NG and NG+ are the grounded VideoQA baselines on NExT-GQA that PSM is built on and compared against."}],"review_version":1}