acceptodds
Withdrawn from ICLR 2027

Observe, Hypothesize, and Verify: Active Evidence-Seeking Reasoning for Training-Free Video Anomaly Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Video Anomaly Detection (VAD) aims to identify rare and complex anomalous events in videos, which requires the model to capture fine-grained details and reason about their contextual relationships. However, existing training-free VAD methods commonly follow a passive caption-and-judge paradigm, where video clips are first converted into generic textual descriptions and anomaly reasoning is restricted to evidence already captured by generic video descriptions via vision-language models. While effective for visually salient events, this paradigm is insufficient for evidence-demanding anomalies, which depend on subtle object interactions, temporally distributed clues, and careful verification. Therefore, we propose OHV-VAD, which formulates video anomaly detection as an observe-hypothesize-verify reasoning process for active evidence seeking and verification. This paradigm is built around three evidence-centered capabilities: 1) Hypothesis-Driven Evidence Acquisition transforms passive captioning into targeted perception by using diagnostic questions as evidence probes to acquire anomaly-relevant visual evidence. 2) Adaptive Temporal Evidence Refinement maintains a compact evidence state through a clue memory that continuously organizes anomaly-relevant clues across video segments. 3) Reflective Anomaly-Normality Verification evaluates each suspicious segment under competing abnormal and normal explanations, mitigating over-detection caused by one-sided anomaly reasoning. Extensive experiments on VAD benchmarks demonstrate strong overall performance among training-free approaches. Category-wise analyses further show that OHV-VAD effectively detects evidence-demanding anomalies while retaining strong performance on visually salient events, and false-positive analyses confirm its ability to reduce over-detection of normal behaviors.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Video Anomaly Detection, Multi-modal LLM

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_TejuOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents OHV‑VAD, a training‑free video anomaly detection framework. It reformulates VAD as an observe‑hypothesize‑verify reasoning pipeline for active evidence collection. Three core components are developed. The method outperforms other training‑free baselines on UCF‑Crime and XD‑Violence. Ablation studies, per‑category analysis and qualitative case visualizations are provided. The idea targets limitations of the existing passive caption‑and‑judge paradigm. However, several key arguments lack solid experimental support. Some important comparisons and real‑world validations are missing. Writing details are not fully polished. This paper does not meet ICLR acceptance bar, and I recommend rejection.

Critical strengths
  1. The paper correctly identifies limitations of the passive caption‑and‑judge paradigm for training‑free VAD, especially for hard anomalies requiring scattered temporal evidence.

  2. The overall framework is conceptually clear. It consists of hypothesis‑driven evidence acquisition, adaptive temporal evidence refinement and reflective anomaly‑normality verification. Qualitative case visualizations illustrate the full reasoning flow.

Critical weaknesses
  1. The comparison against PANDA is superficial. Both belong to agent‑style VAD systems. The paper only reports numeric results. It claims memory design differences in text without supporting quantitative experiments.

  2. The evidence‑probe module heavily relies on manually crafted prompts. Although prompt sensitivity tests are included, the authors do not test its usability under open‑world settings where anomaly categories are unknown in advance.

  3. Failure case analysis is shallow. Only one failure type caused by occlusion and low video quality is shown. The paper does not report statistics for different error modes, nor analyze intrinsic model‑level failure patterns.

  4. Fine‑grained runtime breakdown is missing. The paper does not report separate inference cost for each of the three main modules. Readers cannot identify which component forms the major computational bottleneck.

  5. The temporal evidence memory involves many hyper‑parameters. The maximum memory size of 50 entries is chosen without ablation study to justify this specific value.

  6. OHV‑VAD targets online streaming scenarios. Detailed comparison against other online training‑free VAD baselines is insufficient. Several compared methods are not differentiated by online‑offline settings.

  7. Several qualitative examples show model predictions conflicting with ground‑truth labels. The authors only mention this observation briefly and do not discuss its potential impact on evaluation metrics.

Questions for authors

See the above weakness

Reviewer_tRjiOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors argues that training-free VAD methods such as LAVAD, Monitor and AnomalyRuler follow a passive “caption-and-judge” paradigm, in which anomaly reasoning is limited by the content of the caption the model generates. They propose OHV-VAD, which processes 8-second segments causally with three modules to perform: i) Hypothesis-Driven Evidence Acquisition (HDEA). ii) Storing evidence sentences in a memory and adaptively refining them (ATER module), if they are deemed not to be redundant. iii) An LLM executed Reflective Anomaly–Normality Verification (RANV).

The validation is carried out on UCF-Crime 85.67 AUC; XD-Violence 89.03 AUC / 72.05 AP.

Critical strengths
  1. The key idea is to look for category-specific evidence. This is particularly helpful for theft, shoplifting or burglary which are prevalent in the analysed video sets.
  2. The design uses only past segments.
  3. The component ablations are informative regarding the merits of the proposed three innovations.
Critical weaknesses
  1. The proposed training-free approach relies on knowing the benchmark label space. For fairness, it would be important to compare the performance with a generic “any anomaly” instruction.
  2. It is not clear how “normality” priors are determined.
  3. The impact of the choice of backbone is confounded with the proposed method
  4. The results on shoplifting are lower than SOTA, although this was the motivating example in the introduction.
  5. All ablation are performed only on one dataset.
  6. The individual components have limited novelty. Perhaps they can be considered innovative in the context of VAD, but not in the context of general video understanding.
  7. There are many references that present related work and they are not cited. These include for instance AnyAnomaly (Ahn etal), Flashback (Lee etal 2025), Cerberus (Zheng etal 2025), CUVA(Du etal CVPR 2024), UCA (Yuan et al CVPR 2024) and VAU-R1 (2025)
Questions for authors
  1. What is the performance with a generic instruction (no category list), with a mismatched taxonomy, and for LAVAD or Monitor when given the same 13-category list?
  2. From what data are the normality priors estimated?
  3. What do LAVAD, Monitor and AnomalyRuler achieve using the same Qwen3-VL-8B + Qwen3-14B backbone?
Reviewer_QySuOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes OHV-VAD, a training-free video anomaly detection framework that reformulates conventional passive caption-and-judge approaches into an Observe–Hypothesize–Verify reasoning process. The framework consists of three main components: Hypothesis-Driven Evidence Acquisition (HDEA), which actively seeks anomaly-relevant visual evidence through diagnostic questions; Adaptive Temporal Evidence Refinement (ATER), which maintains a compact memory of informative clues across video segments; and Reflective Anomaly-Normality Verification (RANV), which evaluates suspicious events from both abnormal and normal perspectives. Experiments on UCF-Crime and XD-Violence demonstrate competitive performance against existing training-free methods.

Critical strengths
  1. The paper identifies a meaningful limitation of existing caption-and-judge approaches, where important anomaly-related evidence may be lost during generic video captioning. The proposed active evidence-seeking paradigm provides a reasonable direction for addressing this limitation.
  2. The experiments are relatively extensive, including comparisons with existing methods, ablation studies, category-wise evaluations, and analyses of false positives and computational efficiency. The results demonstrate the effectiveness of the proposed framework.
  3. The paper is generally well written, with a clear motivation and a logical organization of the proposed framework. The methodology and experimental findings are explained in a coherent and understandable manner.
Critical weaknesses
  1. Unclear initialization of the hypothesis-driven reasoning process. According to Eq. (4), the generation of hypotheses and selection of evidence probes depend on previous observations and accumulated evidence memory. However, the observations themselves are generated through hypothesis-guided visual perception in Eq. (5). It is unclear how this process is initialized for the first video segment, when neither historical observations nor evidence memory are available. The authors should clarify how the initial hypothesis and evidence probes are generated, and whether an additional observation step is required to address this cold-start issue.
  2. Dependence on predefined anomaly categories and limited evidence of generalization. The proposed framework relies on anomaly-specific diagnostic evidence priors and question banks. In the experiments, these are constructed using the anomaly categories predefined by UCF-Crime and XD-Violence. This raises concerns about the generalization capability of the framework, particularly in open-domain scenarios where anomaly categories may be unknown or cannot be specified in advance. Although the authors claim that the framework supports user-defined anomaly categories and descriptions, the experiments are conducted using predefined anomaly categories from the benchmark datasets. It remains unclear whether the proposed framework can generalize to previously unseen anomaly types or open-domain anomaly detection scenarios.
  3. The framework figures could better communicate the core methodological ideas. In Figure 1, the right-hand illustration does not adequately reflect the key contributions of the proposed framework, particularly the temporal evidence memory update and the verification process involving competing abnormal and normal explanations. Figure 2 also contains numerous icons and decorative elements that draw attention away from the actual technical details and information flow. The authors are encouraged to simplify the visual presentation and place greater emphasis on the core mechanisms, particularly how evidence is acquired, how memory is updated, and how abnormal and normal explanations are jointly evaluated.
Questions for authors

Please see the "Weaknesses".

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.