{"id":"92d01c62-bb65-4813-8bea-f6e0d950f48b","arxiv_id":"2608.11260","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A training-free glance-then-scrutinize pipeline and a reinforcement-learned tool-calling agent both improve joint temporal grounding and semantic understanding of video anomalies on a new benchmark.","lead":"Surveillance video anomaly detectors usually tell you when something odd happens or what is happening, but rarely both. This paper builds a two-stage system that first scans the whole video to find suspicious moments and then zooms in to explain them, and it also trains an AI agent to do this automatically and in real time.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Agentic RL's reported grounding factor (0.804) implies mean raw IoU of only ~0.1–0.3 under Eqs. 17–18, so the claimed precise temporal grounding is not supported by the paper's own metric.","rationale":"The reader's weakest assumption was that JeAUG is unvalidated and hand-fitted; my concern is more specific and, if it lands, more damaging: even on the authors' own equations, the reported grounding factor for the headline Agentic RL model corresponds to a raw IoU of roughly 0.1–0.3, not the human-consensus 0.7. This means the central claim's temporal-precision component is not merely lacking external validation; it is internally inconsistent with the numerical evidence presented. The concern is load-bearing because the headline result (A.U. 7.35, JeAUG 5.91) is the product of this grounding factor and the understanding score, and the paper explicitly claims the agentic model delivers precise temporal grounding at real-time speed. My proposed check—reporting raw per-video IoU on the held-out set—would settle the issue directly. Since the reader already assigned a CONDITIONAL verdict based on metric concerns, this deeper version of the same concern does not change the verdict category; it strengthens the conditions under which acceptance should be granted, namely release of per-video raw IoU and ground-truth alignments, not just the aggregated JeAUG score.","tokens_in":24697,"tokens_out":8048,"duration_ms":80807,"concrete_test":"Compute the raw per-video IoU (Eq. 15) of the Agentic RL model's final crop window on the 1,217 held-out VAGU-T videos and report its mean, median, and distribution. If the mean raw IoU is below 0.3, the temporal-precision claim in the central result is contradicted; if it is above 0.7, the reported grounding factor 0.804 must be reconciled with Eqs. 17–18 and the actual video-length distribution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the Agentic RL model resolves the \"when-what\" dissociation rests on the grounding half of JeAUG. The paper reports a grounding factor γ·F(IoU) of 0.804 for Agentic RL (Fig. 8). Unpacking Eq. 17 with Eq. 18 at the stated average video length (2,716 frames) gives γ≈1.25, so F(IoU)≈0.64. Because F is piecewise with floor(10·IoU), this F corresponds to floor(10·IoU)=1, i.e., a mean IoU between 0.1 and 0.2; even if some evaluation videos are shorter (γ closer to 1), the implied IoU is at most 0.2–0.3. The paper's own human-consensus target is IoU≈0.7. Moreover, min(γ·F(IoU),1) saturates at IoU≈0.3 for long videos, so the metric cannot reward grounding improvements above 0.3 at all. Thus the JeAUG numbers, including the headline JeAUG 5.91, are compatible with temporal localizations that are far from precise; the floor at 0.5 and the length multiplier hide the low raw IoU. The temporal-precision component of the central claim is therefore not established by the reported evidence, independent of whether JeAUG is externally valid.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses the joint problem of temporal grounding and semantic understanding in video anomaly detection (VAD), which it frames as a \"when-what\" dissociation in existing methods. It proposes two approaches: GtS, a training-free pipeline that uses static and dynamic textual guidance to perform coarse-to-fine anomaly localization and understanding with off-the-shelf VLMs, and a tool-augmented agentic VAD method in which a multimodal LLM learns, via cold-start supervised fine-tuning followed by GRPO-style reinforcement learning, to invoke a video-cropping tool and self-correct its temporal hypotheses. To support training and evaluation, the authors extend their prior VAGU benchmark into VAGU-T, adding chain-of-thought tool-calling traces, and introduce the JeAUG metric, which jointly combines an LLM-judged understanding score with a hand-fitted grounding function of temporal IoU. Experiments on VAGU-T report that GtS substantially improves over direct training-free baselines, and that the Agentic RL model achieves the best overall accuracy (A.U. 7.35, JeAUG 5.91) at 148 FPS.","tokens_in":25051,"tokens_out":7089,"duration_ms":72750,"significance":"If the reported results are taken at face value, the agentic method would be a meaningful advance: a single tool-augmented multimodal model that simultaneously delivers semantic anomaly understanding, temporal grounding, and real-time inference. The paper's strengths include a clearly specified two-stage training pipeline, ablations that support the incremental contribution of each reward term, a large human-validated benchmark with QA pairs and tool-calling traces, and a level of methodological detail that makes the approach reproducible in principle. The significance is, however, conditional on the validity of the author-defined JeAUG metric and on the actual raw temporal-grounding accuracy of the models, neither of which is demonstrated in the current manuscript.","major_comments":[{"comment":"The reported grounding factor of 0.804 for Agentic RL is, under the paper's own metric, consistent with very imprecise temporal localization, not precise grounding. For the stated average video length of 2,716 frames, Eq. (18) gives gamma ≈ 1.25, so the grounding factor 0.804 implies F(IoU) ≈ 0.64. In Eq. (17), this value corresponds to floor(10*IoU) = 1, i.e., a mean raw IoU between 0.1 and 0.2, far below the human-consensus anchor of 0.7 discussed in Sec. 6.2. Even if all evaluation videos were much shorter (gamma close to 1), the implied IoU would be only about 0.2. Moreover, because min(gamma*F(IoU),1) saturates at IoU ≈ 0.3 for long videos, JeAUG cannot reward any temporal-precision improvement above this level. The headline JeAUG 5.91 and the 'precise temporal grounding' claim are therefore not supported by the reported evidence. The authors should report raw IoU (and ideally mAP at IoU thresholds such as 0.3/0.5) separately from the A.U. score, and should re-examine the saturation behavior of JeAUG.","section":"Sec. 6.2, Eq. (17), Eq. (18), Fig. 8, Table 5"},{"comment":"The validity of JeAUG itself is not established. The grounding function F(IoU) and the length-compensation factor gamma are manually designed with constants (0.63, 0.5, 0.25, etc.) that are asserted to match human judgment, but the supporting human study is described only verbally: 10 participants on 50 videos, with the claim that even the lowest pairwise IoU was about 0.7. No agreement statistics, no confidence intervals, and no correlation between JeAUG and human ratings are reported, and there is no external validation against established VAD metrics or independent raters. Since all headline comparisons in Table 5 are reported in JeAUG or A.U., the central claims rest on a metric whose validity is asserted rather than demonstrated. The authors should provide quantitative human-alignment evidence and show that the main conclusions are robust to alternative metric choices (e.g., raw IoU, AUC, or thresholded variants).","section":"Sec. 6.1–6.3"},{"comment":"All headline results come from a single evaluation pass with no error bars, no multiple seeds, and no significance tests. This is particularly important for the claimed superiority of Agentic RL over SFT (A.U. 7.35 vs. 6.62, JeAUG 5.91 vs. 4.48) and over the frame/segment-wise baselines, since RL training is stochastic and FPS measurements depend on implementation. The authors should report mean and standard deviation over at least three independent runs (or otherwise quantify variability), and perform a simple significance test (e.g., paired bootstrap) for the main comparisons in Table 5.","section":"Table 5, Sec. 7.2"},{"comment":"The temporal grounding reward R_time in Eq. (15) is computed as the IoU between the final tool-calling window and the ground-truth interval, and the grounding component of JeAUG in Eq. (16) uses the same ground-truth interval through F(IoU). The RL objective therefore directly optimizes the grounding factor that is later reported in the evaluation metric. This alignment by construction does not by itself invalidate the results, but it means the reported JeAUG gains are exactly what the training procedure was designed to maximize. Independent validation of JeAUG and a separate reporting of raw IoU are necessary to determine whether the gain reflects genuine temporal-precision improvement or only metric-specific saturation.","section":"Sec. 5.4, Eq. (15), Sec. 6, Eq. (16)"}],"minor_comments":[{"comment":"The length-adaptive probability P_multi uses an undefined clip notation and the thresholds L_min and L_max are never given numerical values; please specify them and clarify the formula.","section":"Sec. 3.3, Eq. (1)"},{"comment":"The Savitzky-Golay smoothing formula refers to coefficients q_p and a normalization Q that are not defined in the text; please make the notation explicit or cite the standard filter definition.","section":"Sec. 4.1, Eq. (4)"},{"comment":"Inference FPS is reported without specifying the hardware and precision used for the trained models, whereas the training hardware is listed; this makes the FPS comparisons across methods hard to interpret.","section":"Table 4, Sec. 7.1"},{"comment":"The description of the grounding annotation states that annotators reached consensus 'through IoU-based aggregation,' but no aggregation rule or agreement threshold is given; a brief specification would improve reproducibility.","section":"Sec. 3.3"},{"comment":"The coefficient-of-variation comparison in Table 2 supports the claim that A.U. is more stable than ROUGE/BLEU/METEOR on this data, but the table does not report correlation with human judgments, which is the more relevant validity check for a new metric.","section":"Sec. 6.1, Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a self-contained self-evaluation in which the benchmark, the metric, and the methods all originate from the same group, with no third-party validation. The numerical implication that the reported grounding factor corresponds to a raw IoU of roughly 0.1–0.2 is a serious concern and should be addressed before publication; the authors should be asked to report raw temporal-grounding metrics and demonstrate robustness to the choice of evaluation metric. The novelty relative to the authors' own conference paper should also be checked by the editor, since the present article describes JeAUG, GtS, and VAGU as extensions of [18]."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read on arXiv:2608.11260.\n\nThe paper is a solid engineering contribution with a real new ingredient: adapting tool-augmented RL to query-free video anomaly detection. The VAGU-T benchmark with human-validated labels and tool-calling traces is a useful resource, and the two-stage SFT+RL recipe is clearly described and ablated. If the 148 FPS number holds up, that's an attractive practical profile.\n\nBut the central claim — that the agentic model resolves the when-what dissociation — is not supported by the paper's own metric. The reported grounding factor of 0.804 for Agentic RL, unpacked through Eqs. 17-18 at the stated average video length (2,716 frames), implies a raw IoU around 0.1-0.2, far from the human-consensus 0.7 the metric is supposed to encode. The floor at 0.5 and the γ length multiplier hide this. The paper never reports raw IoU for its own methods; it only gives the transformed factor. So the temporal-precision half of the headline result is unestablished.\n\nThe JeAUG metric itself is also hand-fitted and self-referential: the constants in F(IoU) and γ are tuned to a small internal human study, and the RL reward uses the same IoU ground truth, so benchmark and training objective are aligned by construction. There are no error bars, no independent validation against established VAD metrics, and no external human rating. Those are real limitations, not quibbles.\n\nWhat the paper does well: the GtS pipeline is specified precisely enough to reproduce, the ablations isolate the contribution of each component, and the category-wise analysis is informative. The authors also disclose the GtS hyperparameter tuning on discarded videos, which is honest.\n\nWho should read this: anyone building a VAD benchmark or applying tool-augmented RL to open-ended video tasks. Treat the JeAUG scores as relative signals from the authors' own instrument, not as absolute evidence of practical precision.\n\nRecommendation: worth sending to peer review — the benchmark and method deserve referee time — but the referees should push for raw IoU reporting, external validation of JeAUG, and error bars before acceptance.","headline":"A serious benchmark and a plausible RL recipe, but the temporal-grounding claim collapses under the paper's own metric.","tokens_in":25586,"tokens_out":5141,"would_cite":false,"duration_ms":48352,"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 reports that a single tool-augmented multimodal model can both explain what anomaly happened and localize when it occurred at 148 frames per second, closing the trade-off that previously split video anomaly detection into two…","keywords":["Video Anomaly Detection","Anomaly Understanding","Temporal Grounding","Agentic Reinforcement Learning","Tool-Augmented Reasoning","Training-Free Framework","VAGU-T","JeAUG"],"falsifier":"Take a random subset of VAGU-T, have independent raters mark anomaly intervals and rate anomaly descriptions, and compare their judgments to JeAUG's assumed human-consensus IoU of 0.7 and to its score progression; also recompute the head-to-head comparisons with standard metrics used separately (temporal mAP/IoU and ROUGE/BLEU). If JeAUG does not track the independent raters, or if the agentic model's advantage over frame/segment-wise baselines disappears on the standard metrics, the central claim fails.","tokens_in":24512,"feed_emoji":"🎥","tokens_out":12907,"duration_ms":122634,"temperature":0.7,"pith_summary":"Video anomaly detection is split between systems that know when an anomaly occurs but not what it is, and language-model systems that say what happened but cannot say when. The paper argues this split is a symptom of missing unified reasoning and proposes a global-to-local loop—glance across the whole video, scrutinize suspicious segments, and iterate when the guess is wrong—as the common answer. It realizes the loop twice: a training-free pipeline called Glance then Scrutinize (GtS), which lifts a direct video-QA baseline from a JeAUG score of 2.28 to 4.04, and a tool-augmented agentic model trained by cold-start supervised fine-tuning followed by reinforcement learning, reported at 7.35 on anomaly understanding and 5.91 on JeAUG while running at 148 frames per second. The supporting resources are a new benchmark, VAGU-T, with 7,567 human-validated videos and tool-calling reasoning traces, and a new joint metric, JeAUG, that couples semantic quality with temporal precision. A sympathetic reading is that the when–what dissociation is not an inherent trade-off but a trainable capability gap.","feed_headline":"One model names and times video anomalies at 148 fps","feed_subtitle":"One detector resolves the when–what trade-off, beating exhaustive and training-free baselines on a joint score.","key_machinery":"The machinery is the chain-of-tool-thought itself, a single reasoning loop that replaces exhaustive processing: a think block records the model's temporal hypothesis, a crop-video tool returns densely resampled frames from that window, and the model decides whether to answer or re-enter the loop with a corrected window. That loop is trained end-to-end with a joint reward whose three terms supervise semantic accuracy, output format, and temporal precision, with the temporal term defined as IoU between the last tool-call window and ground truth—a choice that makes the only rewarded localization one the model actually inspected. GtS instantiates the same loop without training by using dual textual guidance and CLIP encoders to build a smoothed temporal anomaly curve, then feeding top segments to VQA and VTG models. The benchmark VAGU-T supplies the traces that make the loop learnable, and the JeAUG metric $\\min(\\gamma F(\\mathrm{IoU}),1)\\cdot \\mathrm{Score}_{\\mathrm{A.U.}}$ is the measurement surface on which all headline comparisons are read.","core_discovery":"The paper's central claim is that the apparent trade-off between temporal grounding and semantic understanding in video anomaly detection dissolves when the detector is built around a global-to-local inspection loop. In the training-free version, the loop is assembled from frozen components: a language model turns captions, category lists, and phrase banks into static and dynamic prompts, CLIP-style encoders score each frame or segment against those prompts to produce a smoothed anomaly curve, and high-scoring windows are handed to question-answering and temporal-grounding models for refinement. In the agentic version, the entire loop lives inside one multimodal large language model: the model looks at globally sampled frames, forms a temporal hypothesis, invokes a native video-cropping tool to inspect that window densely, and either commits to an answer or corrects the window and looks again, with the last tool window serving as the temporal localization. Training proceeds in two stages—cold-start supervised fine-tuning on reverse-engineered tool-calling traces, then reinforcement learning with reward $R = R_{\\mathrm{acc}} + R_{\\mathrm{format}} + R_{\\mathrm{time}}$, where $R_{\\mathrm{time}}$ is the IoU of the final inspected window against the human annotation. On the VAGU-T evaluation set the paper reports that the agentic model reaches the best accuracy among all compared methods (A.U. 7.35, JeAUG 5.91) at 148 FPS, and that the reinforcement-learning stage contributes the bulk of the joint-metric gain over supervised fine-tuning.","pith_inferences":["The same glance-scrutinize-think loop could transfer to other open-ended monitoring problems with no predefined query—wildlife surveillance, industrial fault detection, or security footage review—where the anomaly type and location must both be discovered autonomously.","A testable extension is to vary the maximum number of tool rounds and the density of resampled frames; the paper's per-category results suggest the agentic model's largest advantages would appear on short, subtle events such as shoplifting and traffic violations.","Because GtS is training-free and already improves grounding, it could be used as a noisy annotator or trace generator to bootstrap agentic training in new domains without fresh human annotation; the paper does not itself test this loop.","The grounding function $F(\\mathrm{IoU})$ inside JeAUG could be reused as a dense reward during training rather than only an evaluation score, potentially aligning the optimized objective more directly with the reported metric."],"forward_implications":["If the reported numbers are reproducible, a single roughly 7-billion-parameter multimodal model can perform real-time video anomaly detection with both semantic explanations and temporal grounding, removing the need for exhaustive frame- and segment-wise captioning.","The reinforcement-learning gain over supervised fine-tuning (JeAUG 4.48 to 5.91) implies that outcome-based rewards elicit self-correction and targeted inspection that trace imitation alone does not teach.","The training-free GtS result (same-backbone JeAUG 2.28 to 4.04) implies that coarse textual hypotheses can direct scrutiny well enough to be immediately deployable wherever off-the-shelf VLMs are available.","Because VAGU-T's evaluation videos are held out from both SFT and RL, the claimed advantage of the agentic model is a test-set finding, not a training-set artifact, if the trace-construction pipeline is as described.","The JeAUG operating points (JeAUG at least 3, FPS at least 30) give the field a concrete target for what counts as an acceptable joint anomaly-detection system, allowing different methods to be compared on a single combined scale."],"supporting_citations":[{"why":"Supplies part of the real-world video pool and the most closely related benchmark annotations that VAGU-T builds on for anomaly understanding and grounding.","marker":"[8]"},{"why":"LayVAD is the training-free frame/segment-wise pipeline that GtS is compared against on the accuracy-speed trade-off.","marker":"[13]"},{"why":"VITAL is the tool-augmented video-reasoning model that provides the native video-cropping-tool design the agentic method adapts.","marker":"[16]"},{"why":"LongVT supplies the native tool-calling pipeline pattern and GeminiCoT data used in the SFT/RL training corpus for long-video reasoning.","marker":"[17]"},{"why":"The authors' prior VAGU/GtS conference paper supplies the benchmark, the GtS framework, and the JeAUG metric that this article extends.","marker":"[18]"},{"why":"GRPO is the group-relative policy optimization algorithm used as the RL optimizer after cold-start SFT.","marker":"[47]"},{"why":"UCF-Crime is one of the real-world anomaly datasets integrated into VAGU-T.","marker":"[58]"},{"why":"Gemini-2.5-Pro is one of the multimodal models used to reverse-engineer tool-calling traces from ground-truth annotations.","marker":"[60]"},{"why":"GPT-4o is used as the LLM-as-a-judge for trace filtering and for the accuracy reward in reinforcement learning.","marker":"[61]"}],"fun_headline_variants":["Agentic video anomaly detector resolves when–what at 148 fps","From glance to scrutiny: agentic VAD thinks twice at 148 fps","Training-free loop to agentic reasoning: VAD hits 148 fps","Self-correcting video anomaly model nails both time and meaning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the new JeAUG metric measures what practitioners care about: its grounding curve and video-length compensation were hand-fitted to match a small internal human study, and the paper provides no independent validation against established anomaly metrics or outside raters.","fun_headline_variants_meta":{"raw":{"variants":["Agentic video anomaly detector resolves when–what at 148 fps","From glance to scrutiny: agentic VAD thinks twice at 148 fps","Training-free loop to agentic reasoning: VAD hits 148 fps","Self-correcting video anomaly model nails both time and meaning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00088,"raw_usage":{"total_tokens":3917,"prompt_tokens":1174,"completion_tokens":2743,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":790,"completion_tokens_details":{"reasoning_tokens":2666}},"tokens_in":790,"tokens_out":2743,"duration_ms":21310,"temperature":1.0,"reasoning_tokens":2666,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:37:48.526931+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a random subset of VAGU-T, have independent raters mark anomaly intervals and rate anomaly descriptions, and compare their judgments to JeAUG's assumed human-consensus IoU of 0.7 and to its score progression; also recompute the head-to-head comparisons with standard metrics used separately (temporal mAP/IoU and ROUGE/BLEU). If JeAUG does not track the independent raters, or if the agentic model's advantage over frame/segment-wise baselines disappears on the standard metrics, the central claim fails.","supporting_citations":[{"cited_title":"In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp","cited_arxiv_id":null,"evidence_quote":"Supplies part of the real-world video pool and the most closely related benchmark annotations that VAGU-T builds on for anomaly understanding and grounding."},{"cited_title":"In: Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp","cited_arxiv_id":null,"evidence_quote":"LayVAD is the training-free frame/segment-wise pipeline that GtS is compared against on the accuracy-speed trade-off."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp","cited_arxiv_id":null,"evidence_quote":"VITAL is the tool-augmented video-reasoning model that provides the native video-cropping-tool design the agentic method adapts."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Springer Nature 2021 LATEX template Glance, Scrutinize, and Think31 Computer Vision and Pattern Recognition, pp","cited_arxiv_id":null,"evidence_quote":"LongVT supplies the native tool-calling pipeline pattern and GeminiCoT data used in the SFT/RL training corpus for long-video reasoning."},{"cited_title":"In: Proceedings of the AAAI Conference on Artificial Intelligence, vol","cited_arxiv_id":null,"evidence_quote":"The authors' prior VAGU/GtS conference paper supplies the benchmark, the GtS framework, and the JeAUG metric that this article extends."},{"cited_title":"In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp","cited_arxiv_id":null,"evidence_quote":"UCF-Crime is one of the real-world anomaly datasets integrated into VAGU-T."}],"review_version":1}