{"id":"6aad3c2f-4b23-49f3-819c-d88999dbaa54","arxiv_id":"2506.08493","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"UniCaCLF trains temporal instant features to separate real from forged moments relative to each sample's global context, achieving state-of-the-art temporal forgery localization on five public datasets.","lead":"Researchers propose UniCaCLF, a contrastive learning framework that locates forged segments in videos by comparing each moment to the overall global context of the sample. It reports top results on five public deepfake localization benchmarks while using fewer parameters and faster inference than prior systems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed SOTA gains at AP@IoU=0.95 may stem from the training protocol (DIoU loss, Soft-NMS, unspecified τ) rather than the proposed context-aware contrastive mechanism, because the baseline comparison is not fully controlled.","rationale":"The reader's weakest_assumption focuses on the anomaly-detection premise that forged instants are outliers in the frozen feature space. That is an acknowledged and reasonable assumption for content-driven partial forgery, and the paper's benchmark results are consistent with it. The more pressing and specific concern is that the empirical superiority claim is not yet convincingly isolated from the training protocol. The paper controls pre-trained features across baselines but does not control for loss functions, regression loss, post-processing, or hyperparameters such as τ. Since the improvements concentrate at the strict IoU threshold, they could be explained by DIoU and Soft-NMS. The ablation study (Table VII) shows that CaCL and HAO help within the fixed protocol, but that does not establish that the protocol itself is not the main source of the SOTA gap. The reader already noted the missing code, error bars, and unspecified τ, and assigned CONDITIONAL; my concern reinforces that conditional verdict without changing it. A reproducible, protocol-matched re-evaluation would settle whether the central claim holds. The anomaly-assumption limitation is real but is a boundary of applicability, not an internal inconsistency, so I do not treat it as the primary load-bearing concern.","tokens_in":19933,"tokens_out":8371,"duration_ms":92103,"concrete_test":"Re-run UMMAFormer and ActionFormer on LAV-DF and AV-Deepfake1M using exactly the same training protocol as UniCaCLF (focal loss, DIoU loss with φ1=2, the same optimizer and learning rate, the same number of pyramid levels, Soft-NMS, and a specified temperature τ, e.g., τ=0.1 and τ=1.0), while keeping each baseline's own architecture. If the AP@IoU=0.95 gap narrows by more than 50% or becomes non-significant, the claim that the proposed context-aware contrastive components drive the improvement is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that UniCaCLF 'significantly outperforms' SOTA TFL methods. The largest margins are at AP@IoU=0.95 (e.g., 53.61 vs. 37.22 on LAV-DF), a metric highly sensitive to boundary precision. In Section III-E the authors use DIoU regression loss and Soft-NMS post-processing, and in Section IV.A.3 they set loss weights φ1=2, φ2=0.5 but never specify the temperature τ in Eq. 5. Section IV.A.4 states only that 'unified pre-trained features' were adopted for all models; there is no statement that the same classification/regression losses, loss weights, optimizer schedules, post-processing, or feature-pyramid construction were applied to the baselines. Thus the reported AP@0.95 improvements could be caused by these protocol choices rather than by HAO/ACU/CaCL. Without code, error bars, or a precise protocol description, the comparison is not reproducible and the attribution of the gains to the novelty is not established. This is load-bearing because if the gap closes under a matched protocol, the abstract's central claim would be unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UniCaCLF, a universal context-aware contrastive learning framework for temporal forgery localization (TFL). The framework constructs a context-aware perception (CaP) layer with a heterogeneous activation operation (HAO) and an adaptive context updater (ACU), and adds a supervised context-aware contrastive loss (CaCL) computed sample-by-sample. Experiments are conducted on four benchmark datasets (LAV-DF, AV-Deepfake1M, TVIL, HAD) plus a cross-dataset generalization test on Psynd, reporting AP and AR metrics as well as runtime/parameter counts. The authors claim that UniCaCLF significantly outperforms state-of-the-art TFL and temporal action localization methods, with particularly large gains at AP@IoU=0.95.","tokens_in":20382,"tokens_out":7554,"duration_ms":89108,"significance":"If the reported results are robust, the paper makes a solid empirical contribution: it demonstrates that an anomaly-detection formulation of TFL, combined with intra-sample contrastive learning, yields strong localization accuracy while remaining computationally light (7.32M parameters, 4.56 GFLOPs, 35.08 FPS on TVIL). The ablation study on TVIL shows positive contributions from HAO, ACU, and CaCL, and the HAD-to-Psynd cross-dataset experiment is a useful generalization check. The central average-performance claim is supported by Tables I-VI. However, the current text contains a factual overstatement about the LAV-DF results, and the comparison protocol is not fully controlled, which weakens the attribution of the reported gains to the proposed components. These issues are fixable but need to be addressed before publication.","major_comments":[{"comment":"The text states that UniCaCLF \"can outperform all compared methods under both AP and AR\" on LAV-DF, but Table I contradicts this. At AP@IoU=0.5, UniCaCLF scores 97.81, below MFMS (98.47) and UMMAFormer (98.34); at AP@IoU=0.75, UniCaCLF scores 93.11, below MFMS (94.15) and UMMAFormer (93.54). The correct claim is that UniCaCLF achieves the best average AP and the best AR values, not that it wins every threshold. This overstatement should be corrected in the text and abstract.","section":"Section IV-B.1, Table I"},{"comment":"The comparison protocol is not sufficiently controlled to attribute the reported SOTA gains to the proposed HAO/ACU/CaCL mechanism. The paper states only that \"unified pre-trained features\" were adopted for all models; it does not state whether the baselines were retrained with the same classification/regression losses, loss weights, optimizer schedules, feature-pyramid construction, or post-processing. In particular, UniCaCLF uses focal loss, DIoU regression loss, and Soft-NMS, while the baselines use their own heads and losses. Since AP@IoU=0.95 is highly sensitive to boundary precision, the large margins at AP@0.95 (e.g., 53.61 vs. 37.22 on LAV-DF) could stem from these protocol choices rather than from the proposed context-aware contrastive learning. Please provide matched-protocol comparisons, or ablate the DIoU loss and Soft-NMS on the baseline methods.","section":"Section IV-A.4 and Section III-E"},{"comment":"The temperature hyperparameter tau in the context-aware contrastive loss (Eq. 5) is never specified. Section IV-A.3 gives the number of CaP layers L=6, the loss weights phi_1=2 and phi_2=0.5, and the forged-segment labeling threshold, but tau is omitted. Because CaCL is a core contribution and its behavior depends strongly on tau, this omission prevents exact reproduction and makes it impossible to assess the sensitivity of the proposed method to this key parameter. Please report the value used in all experiments and, if possible, include a sensitivity study.","section":"Section IV-D, Eq. (5), and Section IV-A.3"}],"minor_comments":[{"comment":"Rows 3 and 4 have identical checkmark patterns (both show two checkmarks), so the reader cannot tell whether the third row is HAO+CaCL or HAO+ACU. This makes the marginal contributions of CaCL and ACU unreadable. Please relabel the rows unambiguously.","section":"Table VII"},{"comment":"In the BA-TFD+ row, the AR@10 value appears as \"6364\" instead of \"63.64\" (missing decimal point).","section":"Table II"},{"comment":"The text says \"the fewer the Cap layers, the better the model localization performance,\" but Table VIII shows that 6 and 7 layers give the best AP@0.95 and average AP, respectively. This statement should be revised to match the table.","section":"Section IV-C.2 and Table VIII"},{"comment":"The CaCL loss uses a dot product g·x/tau, while HAO uses negative cosine similarity. Please state explicitly whether the features are L2-normalized before the dot product; if they are, say so, since this affects the scale of the loss and the interpretation of tau.","section":"Eq. (5) and Section III-D"},{"comment":"The caption and text say \"we statistic the average cosine similarity\" but do not describe how the SA, SGP, and CaP features were produced for this figure. Please specify the experimental setup so the figure can be interpreted and reproduced.","section":"Fig. 4"},{"comment":"The exclusion of DiMoDif and Vigo due to lack of released code is reasonable. However, please clarify whether the reported baseline numbers are taken from the original papers or from re-implementations using the unified pre-trained features; this is important for judging the fairness of the comparison.","section":"Section IV-A.4"}],"recommendation":"major_revision","confidential_remarks":"The core experiments are well designed, and the average-performance claim is supported. The main risks are the overstatement in the LAV-DF discussion and the lack of a fully matched comparison protocol, which could affect the attribution of the AP@0.95 gains. The missing temperature tau and the ambiguous ablation table are straightforward to fix. I do not see grounds for rejection, but the revision should be checked carefully before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a serious attempt at temporal forgery localization using a contrastive objective against a global context, and the core idea is coherent: forged segments are treated as outliers in instant-feature space, with genuine instants pulled toward the global context and forged instants pushed away. What is actually new is the sample-by-sample supervised contrastive loss, where positives and negatives come from the same video rather than across a batch, plus the HAO/ACU perception layer. The ablations on TVIL are informative: each component contributes, and the efficiency numbers (#Params, GFLOPs, FPS) are real evidence. Credit is due for adopting unified pre-trained features and for trying to keep the comparison fair.\n\nThat said, the soft spots are real but not fatal. First, the abstract and Section IV-B-1 claim UniCaCLF beats all methods under both AP and AR, but Table I contradicts it: UMMAFormer and MFMS have higher AP@0.5 and AP@0.75. The claim should be \"outperforms on average and at strict IoU,\" not \"under both AP and AR.\" Second, the largest margins are at AP@0.95, and the method uses DIoU regression loss and Soft-NMS without stating whether the baselines received the same treatment. That is a protocol-confound concern, not evidence of dishonesty, but the paper needs to either specify matched training details or soften the attribution. Third, tau in Eq. 5 is never specified, and no code or error bars are provided, so the comparison is not fully reproducible. The cross-dataset generalization result is interesting, though limited to one source-target pair.\n\nOverall, the central direction holds up: on average UniCaCLF is consistently ahead, and the sample-wise contrastive framing is a reasonable explanation. But the textual overclaim and the unspecified baseline protocol need fixing before I would trust the AP@0.95 numbers as evidence for the novelty. If the authors release code or provide matched-protocol comparisons for the baselines with DIoU/Soft-NMS, the case becomes much stronger.\n\nThis paper deserves a serious referee. It is a focused contribution to a niche but practically relevant forensic task, and with a revision that fixes the overclaim, specifies tau and the baseline training protocol, and ideally releases code, it would be acceptable. I would bring it to reading group mainly to debate whether the gains come from the contrastive mechanism or the detection head.","headline":"A plausible and well-ablated contrastive framework for temporal forgery localization, with a solid average-SOTA claim that is undercut by a table contradiction and an unspecified baseline protocol.","tokens_in":20748,"tokens_out":2254,"would_cite":false,"duration_ms":28147,"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":"The paper argues that temporal forgery localization is best treated as intra-sample anomaly detection: a forged instant is one whose features stray from the majority of instants, and it builds a contrastive learning framework around that…","keywords":["temporal forgery localization","contrastive learning","anomaly detection","audio-visual forensics","multimedia forensics","partial forgery","deepfake detection","temporal action localization"],"falsifier":"Take a set of partially tampered videos in which the forged clips are known to be statistically undistinguishable from genuine clips in TSN, BYOL-A, ResNet50, and Wave2vec features (for example, forged-to-genuine cosine similarity to the global context is equal in distribution). If UniCaCLF's localization accuracy then falls to the level of random guessing while a reconstruction-based baseline still localizes them, the central outlier-distance claim is refuted.","tokens_in":19778,"feed_emoji":"🎬","tokens_out":6940,"duration_ms":73256,"temperature":0.7,"pith_summary":"The paper's central claim is that temporal forgery localization should be treated as intra-sample anomaly detection: in a partially tampered video, a forged temporal instant is one whose features stand far from the majority of instants, i.e., from the global context. Building on that premise, the paper proposes UniCaCLF, a contrastive learning framework with a context-aware perception layer and a context-aware contrastive loss, and asserts it outperforms prior TFL and temporal action localization methods on five benchmarks. The largest reported gains appear at the strictest evaluation point, AP@IoU=0.95, where the method reaches 53.61% versus 37.22% for UMMAFormer on LAV-DF, and 52.71% versus 41.17% on AV-Deepfake1M. If correct, this shows that contrasting each instant against a carefully re-estimated global context yields more discriminative features for locating small forged segments than reconstruction- or attention-based approaches.","feed_headline":"Forged video moments are located by contrast with global context","feed_subtitle":"It outscores prior detectors at strict overlap, with the biggest gains on the hardest IoU threshold.","key_machinery":"The load-bearing object is the context-aware perception (CaP) layer together with the context-aware contrastive loss (CaCL). The CaP layer's heterogeneous activation operation (HAO) computes the negative cosine similarity between each instant feature and the global context, applies ReLU to keep only the deviating instants, and re-weights those features so forged instants become more salient; the adaptive context updater (ACU) then forms a cleaned global context by softmax-weighting instants by their similarity to the previous context. CaCL uses that cleaned context as the query in a supervised contrastive objective computed sample-by-sample, pulling genuine instants closer and forged instants farther. These components run inside a multi-scale feature pyramid whose outputs feed classification and boundary-regression heads.","core_discovery":"UniCaCLF discovers forged temporal segments by treating them as outliers relative to the global context of the same input sample, not by comparing across samples. The context-aware perception layer strengthens instant features whose negative cosine similarity to the global context is positive, while an adaptive context updater recomputes the context so forged instants contaminate it as little as possible. A sample-by-sample supervised contrastive loss then pulls genuine instants toward the global context and pushes forged instants away from it. The paper reports that this combination outperforms state-of-the-art competitors across visual-only, audio-only, and audio-visual partial forgery datasets, with particularly large advantages at the tight IoU=0.95 threshold.","pith_inferences":["A testable extension is to apply the same global-context contrast to other sequential forgery domains, such as manipulated text or sensor streams, whenever a frozen encoder produces embeddings in which tampered tokens are outliers.","The components' value could be isolated further by ablating the adaptive context updater against an oracle context computed from ground-truth genuine instants only; this would show how much remaining error comes from context contamination.","The single-global-context assumption may weaken when forged segments occupy a large fraction of the video, since the 'majority' becomes ambiguous; testing on videos with fake-to-real ratios approaching 1:1 would probe that boundary.","Pairing the outlier-based signal with a reconstruction-based consistency signal could cover forgeries that are semantically consistent but not statistically anomalous."],"forward_implications":["At the strictest overlap threshold (AP@IoU=0.95), UniCaCLF beats UMMAFormer by 16.39 points on LAV-DF and 11.54 points on AV-Deepfake1M, and beats TriDet by 6.01 points on the HAD audio-only dataset.","The same architecture and training recipe work for video-only, audio-only, and audio-visual partial forgery by swapping frozen pre-trained encoders, and it generalizes across datasets (HAD to Psynd) better than the compared baselines.","Because the contrastive loss is computed sample-by-sample, performance does not depend on forged segments from different manipulation types being comparable to each other within a batch.","The CaP layer avoids the self-attention rank-loss problem by replacing dense attention with similarity-based activation, which the paper reports makes the model faster and lighter than transformer-based TFL baselines."],"supporting_citations":[{"why":"Supplies the LAV-DF content-driven audio-visual forgery benchmark and the BA-TFD+ baseline that motivated the task.","marker":"[9]"},{"why":"Proposes the UMMAFormer reconstruction-based TFL baseline and the TVIL visual-only dataset used for comparison and ablation.","marker":"[13]"},{"why":"Introduces the TriDet detector and the SGP layer, providing the rank-loss critique and the temporal-action baseline the paper compares against.","marker":"[36]"},{"why":"Documents the rank-loss problem of self-attention, which motivates replacing dense attention with the heterogeneous activation operation.","marker":"[15]"},{"why":"Provides the supervised contrastive learning formulation on which the sample-by-sample context-aware contrastive loss is built.","marker":"[46]"},{"why":"Supplies the pre-trained TSN video feature extractor used for LAV-DF, TVIL, HAD, and Psynd experiments.","marker":"[47]"},{"why":"Supplies the pre-trained BYOL-A audio feature extractor used for the same four datasets.","marker":"[49]"},{"why":"Supplies the pre-trained Wave2vec audio feature extractor used for AV-Deepfake1M.","marker":"[54]"},{"why":"Provides the focal loss used as the classification loss to handle fake-genuine instant imbalance.","marker":"[50]"},{"why":"Provides the distance-IoU loss used for boundary regression of forged segments.","marker":"[51]"}],"fun_headline_variants":["Contrast with global context pinpoints video forgeries","Localizing deepfakes with context-aware contrast","Context-aware contrast finds tampered clips","Temporal forgery localization beats prior at tight IoU","UniCaCLF: context-aware contrast for precise forgery location"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework assumes that a forged temporal segment always behaves as a statistical outlier in the pre-trained feature space, standing noticeably farther from the video's average features than genuine moments do; if tampering is seamless enough to keep forged features inside the normal distribution, the method has no signal to detect it.","fun_headline_variants_meta":{"raw":{"variants":["Contrast with global context pinpoints video forgeries","Localizing deepfakes with context-aware contrast","Context-aware contrast finds tampered clips","Temporal forgery localization beats prior at tight IoU","UniCaCLF: context-aware contrast for precise forgery location"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00102,"raw_usage":{"total_tokens":4295,"prompt_tokens":928,"completion_tokens":3367,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":3290}},"tokens_in":544,"tokens_out":3367,"duration_ms":26283,"temperature":1.0,"reasoning_tokens":3290,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:08:58.232004+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of partially tampered videos in which the forged clips are known to be statistically undistinguishable from genuine clips in TSN, BYOL-A, ResNet50, and Wave2vec features (for example, forged-to-genuine cosine similarity to the global context is equal in distribution). If UniCaCLF's localization accuracy then falls to the level of random guessing while a reconstruction-based baseline still localizes them, the central outlier-distance claim is refuted.","supporting_citations":[{"cited_title":"Glitch in the Matrix: A Large Scale Benchmark for Content Driven Audio-Visual Forgery Detection and Localization","cited_arxiv_id":"2305.01979","evidence_quote":"Supplies the LAV-DF content-driven audio-visual forgery benchmark and the BA-TFD+ baseline that motivated the task."},{"cited_title":"Um- maformer: A universal multimodal-adaptive transformer framework for temporal forgery localization,","cited_arxiv_id":null,"evidence_quote":"Proposes the UMMAFormer reconstruction-based TFL baseline and the TVIL visual-only dataset used for comparison and ablation."},{"cited_title":"Byol for audio: Self-supervised learning for general-purpose audio represen- tation,","cited_arxiv_id":null,"evidence_quote":"Supplies the pre-trained BYOL-A audio feature extractor used for the same four datasets."}],"review_version":1}