{"id":"26450737-80a0-4f6b-a26e-3e11d203ee47","arxiv_id":"2412.01095","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"VERA adapts a frozen vision-language model to video anomaly detection by learning prompt questions through verbalized learning, achieving strong explainable detection on UCF-Crime and XD-Violence.","lead":"A new training trick lets a frozen vision-language model detect video anomalies and explain them, using only coarse video-level labels and no parameter updates. The system, VERA, learns a set of guiding questions by chatting between two copies of the same model, then uses those questions at test time.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The fair-comparison claim rests on unstated baseline conditions: ROC-AUC comparisons against Holmes-VAD and VADor use no-instruction-tuning variants, while the published versions of those methods use instruction tuning or full training, so the SOTA-explainable claim is not yet established on the…","rationale":"The reader identified the baseline-comparison issue as one of three concerns and the visual-perception limitation as the weakest assumption. I agree that the visual-perception ceiling is a real and self-acknowledged limitation, but it is a limitation of the entire VLM-based VAD paradigm, not a flaw specific to VERA's central argument. The paper's own Limitations section flags it, and the method's contribution is precisely to work within that ceiling. The more load-bearing issue for the paper's headline claim is the fairness of the empirical comparison. Section 4.2 says VERA outperforms Holmes-VAD and VADor 'without IT, as reported in their papers,' and the tables footnote the same. The reader's verdict already flags this, so this is partial agreement. I do not see an internal inconsistency in the temperature hyperparameter: the body text in Section B.4 says τ is set to 10, while the sensitivity table shows τ values from 10e-8 to +∞ and says the implementation uses [0.01, 1]. The discrepancy is confusing but the reported default in the main text (τ=10) is what matters and the sensitivity analysis shows the method is robust to τ. The single-run, no-seed issue is secondary because the method's stochasticity is in the VLM decoding and the reported gains are large relative to typical variance, though error bars would still be welcome. The fairness issue, however, directly determines whether the claim 'highest AUC among explainable VAD methods' is true. The AP table on XD-Violence further complicates the fairness story: Holmes-VAD with IT and full training reaches 84.96 AP versus VERA's 70.54. The paper's explanation is that full-frame training helps, which is plausible, but it means the headline comparison mixes training-data usage with method design. The cleanest fix is to run the published IT versions of the baselines and report the numbers in the same table. Until then, the correct claim is: VERA achieves the highest AUC among explainable VAD methods that do not use instruction tuning, with strong but not fully established evidence against the best published versions of prior systems. That is a CONDITIONAL result, not a rejection.","tokens_in":24758,"tokens_out":2088,"duration_ms":16185,"concrete_test":"Obtain the official Holmes-VAD and VADor checkpoints that include instruction tuning, run them on the UCF-Crime and XD-Violence test splits under the same protocol (same frame sampling, same segment definitions, same AUC computation), and compare against VERA's reported 86.55 and 88.26. If either baseline exceeds VERA, the SOTA-explainable claim in Section 4.2 must be revised to a conditional claim. Also re-run Table 2 with the Holmes-VAD and VADor IT numbers included in the explainable block.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central empirical claim in Section 4.2 is that VERA achieves the highest AUC among explainable VAD methods on UCF-Crime, outperforming Holmes-VAD and VADor in a fair comparison. The phrase 'without IT, as reported in their papers' is the load-bearing qualifier. The published Holmes-VAD and VADor are instruction-tuned systems, and their instruction-tuned versions report higher AUC (e.g., Holmes-VAD and VADor with IT exceed 86.55 on UCF-Crime). Comparing only to no-IT versions conflates two variables: the method's architecture and the training regime. VERA's whole selling point is that it avoids instruction tuning, so the comparison should be against the best published version of each baseline, with a clear statement of what IT adds. Additionally, the XD-Violence AUC table (Table 3) omits Holmes-VAD entirely, and the AP table (Table 10) shows VERA at 70.54 versus Holmes-VAD at 84.96, a gap the paper explains by training-frame usage but does not reconcile with the fairness claim. The concern is not that VERA is weak; it is that the headline SOTA claim is underdetermined by the reported comparisons. A concrete test: run the published Holmes-VAD and VADor checkpoints (with IT) on UCF-Crime and XD-Violence under the same evaluation protocol and report AUC alongside VERA. If VERA still leads, the claim stands; if not, the claim should be softened to 'best among methods that do not use instruction tuning or full-frame training.'","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VERA, a verbalized-learning framework that adapts a frozen vision-language model to weakly supervised video anomaly detection. The method treats a small set of natural-language \"guiding questions\" as learnable parameters, refining them through a learner-optimizer loop in which both agents are implemented with the same frozen VLM. At inference, the learned questions are inserted into the prompt, segment-level anomaly scores are generated, and these scores are refined by scene-context retrieval (ImageBind) and temporal Gaussian/position weighting. Experiments on UCF-Crime and XD-Violence report frame-level AUC of 86.55 and 88.26, respectively, together with ablations, transfer experiments, and qualitative examples of explainable predictions.","tokens_in":25009,"tokens_out":4740,"duration_ms":44359,"significance":"If the empirical claims hold, VERA is a meaningful contribution: it demonstrates that a frozen VLM can be adapted to VAD without parameter updates, instruction tuning, or an external reasoning LLM, using only coarse video-level labels, while also providing textual explanations. The training algorithm is described transparently in Algorithm 1, and the ablation study in Section 4.3 gives useful evidence about which components matter (question learning, scene retrieval, temporal refinement). The transfer tables (Tables 7-9) and the sensitivity analysis in Appendix B.4 are informative and suggest the method is not brittle in every hyperparameter. However, the headline state-of-the-art claim is currently underdetermined by the baseline comparison used and by the absence of repeated-run statistics.","major_comments":[{"comment":"The claim that VERA achieves the highest AUC among explainable VAD methods on UCF-Crime rests on comparing against Holmes-VAD and VADor \"without IT, as reported in their papers.\" The published versions of these methods are instruction-tuned systems, and their instruction-tuned variants report higher AUC in their own papers. This comparison conflates the proposed architecture with the absence of instruction tuning. The headline claim should be supported either by evaluating the published instruction-tuned checkpoints under the same protocol, or by explicitly softening the claim to \"best among explainable methods that do not use instruction tuning or full-frame training.\"","section":"Section 4.2, Table 2"},{"comment":"The XD-Violence evaluation is incomplete with respect to the same state-of-the-art claim. Holmes-VAD is omitted from the AUC table (Table 3), while in the AP table (Table 10) Holmes-VAD outperforms VERA by a large margin (84.96 vs 70.54). The AP gap is attributed to full-frame training, but then the \"fair comparison\" in Table 2 should also apply to the XD-Violence AUC, where the missing Holmes-VAD row makes the claim untestable. Report Holmes-VAD AUC on XD-Violence under both its published and no-IT configurations, with a clear statement of which training frames each configuration uses.","section":"Tables 3 and 10"},{"comment":"All experiments are single-run: no seeds, standard deviations, or significance tests are reported for any table. Because training involves random batch sampling and stochastic VLM decoding, the point differences supporting the main claim (e.g., 86.55 vs 85.90 for VADor on UCF-Crime) may lie within run-to-run noise. Please report results over multiple seeds for the main tables, or otherwise provide a statistical significance analysis, and state how Q* and the inference hyperparameters are selected without test-set feedback.","section":"Section 4 and Appendix B"},{"comment":"Several inference hyperparameters (K, tau, omega, sigma_1, sigma_2) are set per dataset, and the sensitivity analysis is performed only on UCF-Crime. The final AUC is visibly sensitive to some of these choices (Table 13 shows 85.21 without retrieval and 86.61 at K=0.15h). It is therefore important to specify how these values are chosen for XD-Violence and for future datasets. If any of these choices are made by looking at test-set performance, the reported numbers are optimistic; if not, the validation protocol should be described explicitly.","section":"Appendix B.4"}],"minor_comments":[{"comment":"The caption says \"No IT is used for Holmes-VAD\" but the superscript dagger is only explained after the table; please clarify in the caption which methods use full training frames and which use no instruction tuning, since this distinction is central to the paper's claims.","section":"Table 10 caption"},{"comment":"There is a grammatical error in the sentence \"Additionally, We also evaluate the transferability of Q across different scales and and observe an interesting phenomenon\": the doubled \"and\" should be removed.","section":"Section 4.3, Table 7"},{"comment":"The concluding sentence claims VERA works \"without additional computation overhead,\" but inference uses an ImageBind feature extractor and convolutional smoothing on top of the VLM forward passes. Please rephrase to \"without additional reasoning modules or parameter updates\" to avoid overstating the computational claim.","section":"Section 5"},{"comment":"The phrase \"state-of-the-art explainable VAD performance\" appears in the abstract and contributions before the limitations of the comparison are acknowledged; please qualify it in the abstract to match the actual experimental setup.","section":"Abstract and Introduction"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is likely to be a solid contribution after the baseline comparison is repaired. The authors should either run the published instruction-tuned versions of Holmes-VAD and VADor under the same evaluation protocol, or recalibrate all SOTA language to the no-IT/no-full-frame-training regime. A code release would also substantially increase the value of the paper, since the project page does not appear to provide runnable artifacts."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Thanks for sharing the VERA paper. My take: the core idea is genuinely new and worth engaging with, but the empirical claims are overreached in the current form.\n\nWhat's new and good: VERA is the first application of verbalized machine learning to video anomaly detection. It learns a set of prompt questions for a frozen VLM from coarse video-level labels, with a learner/optimizer loop, avoiding instruction tuning. The ablations are thorough: sampling strategies, question number, inference steps, hyperparameter sensitivity, and transfer across backbones and datasets. The qualitative examples show real explainability benefits, and the method preserves the VLM's general question-answering ability, which is a fair point against IT models. The limitations section honestly flags the dependence on the VLM's visual encoder.\n\nSoft spots are real but addressable. First, the headline SOTA claim in Sec 4.2 compares against Holmes-VAD and VADor without instruction tuning, but the published versions of those methods use IT and report higher numbers. The claim \"highest AUC among explainable VAD methods\" is underdetermined; at best it's \"best among methods without IT or full-frame training.\" The XD-Violence AP table shows a large gap vs Holmes-VAD (70.54 vs 84.96), explained by training-frame usage, but not reconciled with the fairness claim. Second, the experiments appear to be single-run, with no seeds, no error bars, and no code release. For a stochastic VLM-based optimization loop, that's a real gap. Third, there's a minor internal inconsistency: the default temperature tau is set to 10, but the sensitivity discussion recommends choosing tau in [0.01, 1].\n\nNone of this kills the method. The idea is plausible, and the ablations suggest the learned questions matter. But the paper would be much stronger with multiple seeds, released code, and a fair comparison against the best published baselines, either by running the IT checkpoints or clearly limiting the claim.\n\nWho's this for? Researchers in VAD and in verbalized learning/prompt optimization. It deserves a serious referee, not a desk reject. I'd recommend peer review with major revision.","headline":"VERA applies verbalized learning to VAD with a genuinely novel idea, but the SOTA claim rests on unfair baselines and single-run evidence.","tokens_in":25654,"tokens_out":3210,"would_cite":true,"duration_ms":27491,"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":"VERA claims that a frozen vision-language model can perform explainable video anomaly detection by learning a small set of natural-language guiding questions through a verbal learner–optimizer loop, using only coarse video-level labels…","keywords":["video anomaly detection","explainable anomaly detection","vision-language models","verbalized learning","learnable prompts","weakly supervised learning","frozen models","prompt optimization"],"falsifier":"Run VERA on UCF-Crime with the optimizer replaced by a non-learning prompt shuffler that paraphrases the current questions without seeing frames or labels; if the paraphrased prompts reach the same 86.55% AUC, the learner–optimizer verbal loop is not what carries the result, and if they fall well short, the data-driven optimization is load-bearing.","tokens_in":24417,"feed_emoji":"🎥","tokens_out":13092,"duration_ms":108111,"temperature":0.7,"pith_summary":"This paper tries to establish that a frozen vision-language model can integrate description and reasoning for video anomaly detection without instruction tuning or an external reasoning module. The proposed method, VERA, learns a small set of natural-language guiding questions that decompose the vague notion of \"anomaly\" into concrete, checkable patterns such as people out of place, vehicles misused, or visible damage. These questions are optimized through verbal feedback between a learner VLM and an optimizer VLM, using only coarse video-level labels, and are then embedded in prompts at inference time to produce segment-level and frame-level anomaly scores. If true, explainable anomaly detection becomes much cheaper: no parameter updates, no frame-level annotation, and no extra LLM calls during inference.","feed_headline":"Frozen VLM beats explainable baselines by learning its own questions","feed_subtitle":"Trained only on coarse labels, learned text prompts beat prior explainable detectors without weight updates.","key_machinery":"The central object is the learned guiding-question set $Q$, treated as language-based learnable parameters. A learner agent, the frozen VLM prompted with $Q$, makes a binary video-level prediction on uniformly sampled frames; an optimizer agent, the same frozen VLM given a different prompt plus batched frames, predictions, and ground-truth labels, rewrites $Q$. At inference, the best $Q^*$ is inserted into the prompt to score 10-second segments, and those scores are refined by ensembling with scene-similar segments using cosine similarity and ImageBind features, then by Gaussian smoothing and position weighting to yield frame-level scores. The questions carry the argument: they operationalize \"anomaly\" into patterns the VLM can verify.","core_discovery":"The paper's central claim is that the bottleneck for frozen VLMs in video anomaly detection is the prompt, not the model. With questions that describe concrete abnormal patterns, a frozen InternVL2-8B reaches 86.55% AUC on UCF-Crime and 88.26% on XD-Violence, reported as the best among explainable methods in a fair comparison, while still being able to answer open-ended follow-up questions. The paper argues that the complex reasoning VAD seems to require can be decomposed into verbal reflections that are learnable, transferable across datasets and models, and sufficient to replace both instruction tuning and external LLM reasoning modules.","pith_inferences":["Beyond the paper, the learner–optimizer verbal loop is a general recipe: any video-understanding task with coarse labels and a checkable visual predicate could be recast as learning a small set of guiding questions, not just anomaly detection.","Beyond the paper, the reported transfer asymmetry, in which questions from the smaller model help the larger one but not vice versa, suggests question complexity should be matched to VLM reasoning ability; one could test this by measuring AUC against the syntactic complexity of learned questions.","Beyond the paper, a stress test follows from the authors' stated limitation: evaluate VERA on anomalies that CLIP-style encoders are known to miss, such as small-object or fine-grained manipulation anomalies; failure there would confirm the perception bottleneck rather than the question-optimization component.","Beyond the paper, the framing suggests a cheap ablation: re-derive the questions with a stronger visual encoder while keeping the same language model, which would localize how much of VERA's lift comes from perception versus verbal prompting."],"forward_implications":["Explainable VAD can run end-to-end with a frozen VLM, so deployment needs no GPU fine-tuning and keeps the model's general question-answering ability intact.","The learned questions transfer across model sizes, architectures, and datasets, so anomaly-detection knowledge can be shipped as text rather than as weights.","Coarse video-level labels are sufficient for training, avoiding the frame-level instruction-tuning datasets used by prior explainable pipelines.","Anomaly scoring can be made context-aware in three cheap steps, segment prompting, scene-similarity ensembling, and temporal smoothing, each of which adds measurable AUC.","Because the backbone stays frozen, VERA can be upgraded simply by swapping in a newer frozen VLM and reusing or re-learning the questions."],"supporting_citations":[{"why":"Supplies the verbalized learner–optimizer procedure that VERA adapts from static tasks to video.","marker":"[45]"},{"why":"Provides the frozen-VLM prompting baseline and the segment-sampling plus scene-retrieval scoring steps VERA builds on.","marker":"[52]"},{"why":"Is the instruction-tuned explainable VAD baseline that VERA must outperform in the UCF-Crime comparison.","marker":"[55]"},{"why":"Is the external-LLM explainable VAD baseline compared without instruction tuning.","marker":"[26]"},{"why":"Supplies the UCF-Crime benchmark, its coarse labels, and the weakly supervised setting used for training and testing.","marker":"[32]"},{"why":"Supplies the XD-Violence benchmark and its coarse labels for the second evaluation setting.","marker":"[42]"},{"why":"Is the frozen InternVL2 vision-language model used as the default learner and optimizer backbone.","marker":"[7]"},{"why":"Provides the ImageBind feature extractor used to retrieve scene-similar segments for score ensembling.","marker":"[10]"},{"why":"Identifies the CLIP-style visual encoder whose fine-grained perception limits VERA, per the paper's limitations.","marker":"[30]"}],"fun_headline_variants":["Learnable text questions unlock frozen VLMs for video anomaly detection","VERA: Frozen VLMs detect anomalies via learned self-questions","Prompt learning beats tuning for explainable VAD","No tuning needed: VLMs learn prompts for anomaly explanation","Explainable VAD without weight updates: VLMs learn their own questions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise, stated in the paper's limitations, is that the frozen VLM's visual encoder perceives the fine-grained details needed to distinguish anomaly from normal; if the encoder misses those cues, the learned questions and the frame-level scores inherit the same blind spots.","fun_headline_variants_meta":{"raw":{"variants":["Learnable text questions unlock frozen VLMs for video anomaly detection","VERA: Frozen VLMs detect anomalies via learned self-questions","Prompt learning beats tuning for explainable VAD","No tuning needed: VLMs learn prompts for anomaly explanation","Explainable VAD without weight updates: VLMs learn their own questions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001071,"raw_usage":{"total_tokens":4474,"prompt_tokens":922,"completion_tokens":3552,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":3467}},"tokens_in":538,"tokens_out":3552,"duration_ms":22671,"temperature":1.0,"reasoning_tokens":3467,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:41:25.761563+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run VERA on UCF-Crime with the optimizer replaced by a non-learning prompt shuffler that paraphrases the current questions without seeing frames or labels; if the paraphrased prompts reach the same 86.55% AUC, the learner–optimizer verbal loop is not what carries the result, and if they fall well short, the data-driven optimization is load-bearing.","supporting_citations":[{"cited_title":"Harnessing large language mod- els for training-free video anomaly detection","cited_arxiv_id":null,"evidence_quote":"Provides the frozen-VLM prompting baseline and the segment-sampling plus scene-retrieval scoring steps VERA builds on."},{"cited_title":"Real-world anomaly detection in surveillance videos","cited_arxiv_id":null,"evidence_quote":"Supplies the UCF-Crime benchmark, its coarse labels, and the weakly supervised setting used for training and testing."},{"cited_title":"Not only look, but also listen: Learning multimodal violence detection under weak supervision","cited_arxiv_id":null,"evidence_quote":"Supplies the XD-Violence benchmark and its coarse labels for the second evaluation setting."},{"cited_title":"Imagebind: One embedding space to bind them all","cited_arxiv_id":null,"evidence_quote":"Provides the ImageBind feature extractor used to retrieve scene-similar segments for score ensembling."},{"cited_title":"Learn- ing transferable visual models from natural language super- vision","cited_arxiv_id":null,"evidence_quote":"Identifies the CLIP-style visual encoder whose fine-grained perception limits VERA, per the paper's limitations."}],"review_version":1}