acceptodds
Under review as a conference paper at ICLR 2027

PARSEE-VAD: Efficient Training-Free Online Video Anomaly Detection via Proposition-Aware Reasoning and Streaming Evidence Escalation

Abstract

Training-free online video anomaly detection (VAD) with frozen multimodal language models faces two coupled challenges: extracting reliable current-window semantics under causal and computational constraints, and maintaining temporal continuity without repeatedly transmitting high-dimensional history. Encoding history through text can compress visual evidence and introduce semantic bias, whereas retaining visual history expands multimodal context. We introduce PARSEE-VAD, a two-module framework that separates semantic evidence acquisition from score-state evolution. Proposition-Aware Reasoning (PAR) extracts structured propositional evidence from the current causal window and conditionally activates more specific queries when coarse evidence warrants further refinement. By sharing a reusable causal visual prefix across queries, PAR reduces redundant computation through selective execution. Streaming Evidence Escalation (SEE) maps the acquired proposition evidence into a compact score-domain event state through current evidence escalation, then propagates only the resulting bounded state across decisions to support temporal continuity. Experiments on four benchmarks demonstrate strong training-free online performance while selective routing reduces specialist computation and score-state propagation remains sparse. These results support a current-first principle for streaming multimodal inference: resolve present semantics first, then use compact historical state only to repair residual continuity gaps.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
video anomaly detection, online video anomaly detection, training-free inference, multimodal language models, proposition-aware reasoning, conditional computation, streaming inference

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_2pgpOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This work proposes PARSEE‑VAD a training‑free online video anomaly detection framework with two modules. PAR extracts structured propositional evidence with shared visual prefix and conditional routing. SEE maintains compact score state instead of high dimensional history. Strengths include clear problem formulation careful efficiency measurement and comprehensive appendix ablation. Main flaws are modest inconsistent benchmark gains inherent structural bias from proposition routing over stated novelty of SEE and incomplete baseline comparison. I recommend reject as overall advancement is incremental and key real world defects remain unaddressed.

Critical strengths
  1. This paper identifies two real pain points for frozen model online VAD scoring interface bias and history transmission bottleneck. Separating evidence acquisition and score state evolution forms clean conceptual design.

  2. Shared visual prefix reuse combined with conditional PA‑Routing cuts computation cost. Latency prefix fidelity and routing preservation receive quantitative measurement.

Critical weaknesses
  1. Empirical improvements against state of the art training free online baselines are inconsistent and limited. Flashback achieves better results on UCF‑Crime, and online SphereVAD performs better on UBnormal. Several performance gains have overlapping bootstrap confidence intervals, which may come from random variance.

  2. Proposition hierarchy contains inherent structural bias. Q3 physical dynamics check controls access to specialist probes P3 and P4. Persistent static hazards such as smoldering fire and stationary damage face a double penalty. Q3 suppresses anomaly score and blocks specialist queries. Main experiments provide no mitigation variant or category-wise quantitative evaluation.

  3. SEE temporal escalation delivers marginal performance benefit. Specifically, the temporal gate activates within only one to three percent of total windows, and bootstrap analysis shows its overall improvement is tiny and statistically insignificant on MSAD. Main performance gain comes from the PAR module.

  4. Baseline comparisons are incomplete in this paper. Specifically, MACD is a strong online training-free baseline, but this paper lacks its XD‑Violence and MSAD experimental results. Also, the paper does not include controlled comparisons of all competing methods under a unified decision cadence and hardware.

  5. The authors should analyze why obvious improvement only appears on XD‑Violence while the proposed method falls behind existing approaches on UCF‑Crime and UBnormal.

  6. The authors should provide complete MACD results on XD‑Violence and MSAD under a matched evaluation setting. Can you give controlled comparison for all online baselines with an identical decision stride?

Questions for authors

See the above weakness

Reviewer_Y3qFOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

PARSEE-VAD investigates how a frozen multimodal model can be used for reliable and efficient online video anomaly detection. Its central idea is to separate semantic evidence acquisition from updates to the score state. PAR reuses the visual prefix of the current window and acquires evidence through proposition-margin readout and conditional querying. SEE integrates the current evidence and uses a small amount of score history to compensate for brief reductions in evidence strength. The paper evaluates detection performance on four benchmarks and uses ablations, budget-matched comparisons, and runtime analysis to validate the role of each component. The results show that the main accuracy gains come from current evidence processing, while temporal maintenance provides smaller additional improvements; prefix reuse and selective querying reduce inference overhead.

Critical strengths
  1. Clear empirical gains from current evidence processing. In comparisons using the same model, the Q2 margin readout outperforms scalar score generation, and current specialist evidence also improves some metrics. The paper offers a practical way to improve performance without updating model weights, which constitutes the main value of the work.

  2. The computational savings are supported by targeted comparisons. Both prefix reuse and on-demand querying are backed by direct runtime measurements. The budget-matched and routing-mask controls also distinguish evidence selection from computational savings. This gives the efficiency improvements an empirical basis and addresses the resource constraints of continuous video processing.

Critical weaknesses
  1. An ongoing fire, smoke, or a person lying motionless after a fall can still indicate danger even when little visible change occurs. Yet even when Q2 has judged the observation to be anomalous, a negative Q3 still causes the subsequent rules to lower the current evidence score and block specialist queries. This bias comes from the method’s design and limits its suitability for persistent hazards. The authors acknowledge the issue, but the existing analysis does not adequately quantify the associated cost in recall. Improvements in aggregate metrics are therefore insufficient to show that this prior is suitable for the anomaly types the method aims to cover.

  2. The main improvement on UCF comes from the Q2 readout, while MSAD AP benefits substantially from current specialist evidence. The overall gains from temporal maintenance are relatively small. In the temporal comparisons, EMA performs better on UCF and XD, whereas the proposed method performs better on MSAD. These results show that performance depends on the dataset, but do not explain which problems the specific history gate addresses that simple smoothing handles poorly. Even as an auxiliary module, its additional value still needs to be demonstrated through results on detection continuity, error persistence, and related outcomes.

  3. The history branch intervenes only under specific conditions, and its low activation rate is related to the gating design. This does not directly establish that the task requires only a small amount of score history. MA, EMA, and Max compare how existing numerical states are processed; they do not compare what information should be retained from the past. Without comparisons against textual or visual history using the same model and comparable budgets, the experiments cannot determine whether richer history still offers additional value. The results therefore support the effectiveness of this particular implementation, but are not yet sufficient to support a general conclusion about how historical information should be used.

Questions for authors
  1. For persistent-hazard categories such as Fire and Arson, how do recall and false alarms change relative to the Q2 baseline after introducing the dynamics gate? How do these category-level results support the rationale for the current gating design?

  2. With the same current evidence, are there results on detection interruptions or persistent false alarms that demonstrate the additional value of the proposed history gate over EMA?

  3. Are there comparisons with textual or visual history using the same backbone and comparable budgets that support the “current-first” claim? If not, how should the scope of this claim be limited?

Reviewer_2Fh1OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes PARSEE-VAD, a training-free framework for online video anomaly detection that separates semantic evidence acquisition from score-state evolution. It consists of two modules: Proposition-Aware Reasoning (PAR) extracts structured propositional evidence from the current causal observation window and conditionally activates more specific queries when needed, while Streaming Evidence Escalation (SEE) integrates the acquired evidence into anomaly scores and propagates compact score states across decisions to maintain temporal continuity. Experiments on four benchmarks demonstrate competitive detection performance with reduced computational overhead.

Critical strengths
  1. Reading a signed logit margin from a binary Normal/Abnormal prompt gives a clear gain in performance.
  2. Detailed empirical reporting.
  3. Honest failure analysis.
Critical weaknesses
  1. The main gain in performance comes from the margin readouts and the use of strong backbone. The proposed refinement methodology (Proposition-Aware Reasoning and Streaming Evidence Escalation) add very little and this should be reflected in the abstract and claims.
  2. The choice of backbone makes a comparison with other approaches misleading, as they use weaker models.
  3. The proposed approach involves many parameters and uses benchmark specific prompts.
  4. Under the anchor-aligned causal view in Table 12, which does not use retrospective scores, the performance drops.
  5. The design appears to have inductive bias which is manifest in suppressing persistent hazards such as fire, smoke, fallen person.
  6. The clarity of presentation is marred by technospeak such as “evidence escalation”.
  7. Relevant papers are missing, e.g VQAScore (Lin et al ECCV 2024) which is very similar to the proposed margin readout idea.
Questions for authors
  1. How were the prompts, gates and constants chosen? Was any benchmark test data, including UCF-Crime and MSAD, used during design? Can you report a leave-one-dataset-out protocol?
  2. What does a baseline re-implemented on Qwen3.5-9B achieve (e.g. LAVAD-style captioning-then-scoring, or Flashback)? What does the Q2 margin readout achieve on the backbones used by the prior methods?
  3. Can you report all headline numbers, for your method and the online baselines, under the anchor-aligned causal mapping, with each baseline’s native mapping stated?
  4. Why does a causal EMA beat SEE on UCF-Crime and XD-Violence (Table 17)? What does SEE offer beyond a well-tuned EMA?
  5. What are the per-category AUCs, especially for persistent anomalies (fire, arson, explosion, fallen person), for Q2-only vs. full PARSEE-VAD?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.