acceptodds
Under review as a conference paper at ICLR 2027

ARIA: Adaptive Reasoning for Video Anomaly Detection with Vision-Language Model

Abstract

Vision-language models (VLMs) have demonstrated strong reasoning capabilities for video anomaly detection (VAD), enabling both anomaly prediction and human-understandable explanation. Existing VLM-based methods typically process videos segment by segment, allocating similar computation to each segment despite their varying needs for further examination. While some segments can be reliably handled with sparse visual observations, others may benefit from additional evidence, either by examining denser visual observations or by incorporating a complementary interpretation from another VLM. Uniformly applying such additional computation to all segments, however, can incur substantial inference cost. To address this issue, we propose ARIA (Adaptive ReasonIng for Video Anomaly Detection), an adaptive reasoning framework that selectively allocates additional computation to video segments when needed. Given a segment, ARIA first assesses the reliability of its initial lightweight VLM response using response-level signals together with temporal and semantic context. For segments warranting further examination, ARIA then determines whether to acquire richer visual evidence through denser frame sampling or introduce complementary textual evidence from another VLM. Experiments on UCF-Crime and XD-Violence demonstrate that ARIA improves anomaly detection and explanation quality while reducing inference cost, yielding a favorable performance-computation trade-off.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Video Anomaly Detection, Vision-Language Models, Adaptive Computation, Collaborative Reasoning, Efficient Inference

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

This paper introduces that there is a significant limitations among existing VLM-based pipelines for VAD where not all the video clip share the same importance contributed to the VAD task.

This paper proposes a fundamental challenge: **How can we enable VLMs to adaptively allocate computation across video segments, allowing more careful examination when needed?""

In solving this problem, this paper proposes ARIA, an adaptive reasoning framework that enhances VLMs for VAD. For segments that warrant additional examination, ARIA determines whether the uncertainty is better addressed by acquiring more visual evidence or by introducing a complementary model perspective.

Extensive experiments on UCF-Crime and XD-Violence show that ARIA improves detection and explanation quality while reducing inference cost.

Critical strengths
  1. The idea is neat and reasonable. The aim of this paper proposes a key insight in VAD research area. The performance over the UCF-crime and XD-violence indicates part of the remarkable effectiveness of the proposed method.

  2. The Ablation of adaptive reasoning components evaluate the effectivenss of the idea.

Critical weaknesses
  1. A lot of baselines are missing such as [1] and [2] limiting the effectiveness of this paper. The number of frames per segment is not specificly organized and number of frames is not tested during the the ablation.

  2. This paper indicating that the utlization of detailed frames may contributes to the better performance over VAD. However, there is a section required for systematically evaluate this phenomenon. It is possible to have a naive experiment on it?

  3. The evaluated benchmarks should not just limited in UCF-crime and XD-violence.

[1]Zou S, Tian X, Wesemann L, et al. Unlocking vision-language models for video anomaly detection via fine-grained prompting[C]//2026 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE, 2026: 4223-4233. [2]Huang C, Wang B, Wang W, et al. Vad-r1: Towards video anomaly reasoning via perception-to-cognition chain-of-thought[J]. Advances in neural information processing systems, 2026, 38: 118486-118518.

Questions for authors

See above.

6 comments

Thank you for the valuable review. Our responses are addressed below.

Comment 1: Can the authors add missing baselines and evaluate the impact of different frame counts?

Answer: Thank you for this suggestion. We added comparisons with ASK-HINT [1] and Vad-R1 [2], clarified the number of frames used by each inference action, and evaluated how frame allocation affects detection performance and inference cost.

(1) Comparisons with ASK-HINT and Vad-R1.

For ASK-HINT [1], we could not find official code and therefore use the result reported in its paper, 87.42% AUC on UCF-Crime with InternVL2.5-8B. ARIA achieves 88.32%, which is 0.90 points higher. The two methods use their own models and inference configurations, so this is a comparison under their respective settings and not a matched-budget comparison.

Vad-R1 [2] focuses on Video Anomaly Reasoning. It uses Perception-to-Cognition Chain-of-Thought for structured anomaly reasoning and combines SFT with AVA-GRPO reinforcement learning. ARIA keeps both VLMs frozen and trains only a lightweight router.

Table 9. Comparison on Vad-Reasoning.

MethodBLEU-2 ↑METEOR ↑ROUGE-2 ↑Acc. (%) ↑F1 (%) ↑mIoU (%) ↑[email protected] (%) ↑[email protected] (%) ↑
InternVL3-8B0.1240.2860.11677.975.655.061.354.0
VideoLLaMA3-7B0.0660.2000.09266.562.442.545.141.9
Qwen2.5-VL-7B0.1130.2640.11676.173.056.761.056.3
VideoChat-R1-7B0.1280.2870.12379.379.055.964.254.0
GPT-4o0.1540.3410.13371.176.047.256.547.6
Vad-R1-7B [2]0.2330.4060.19487.586.271.377.070.6
ARIA (ours)0.1460.3680.20685.483.053.459.855.3

Compared with its Base VLM, ARIA improves accuracy by 7.5 points and F1 by 7.4 points, and both scores exceed all listed baselines except Vad-R1. All three description metrics also improve over InternVL3-8B. For temporal localization, ARIA's [email protected] is 1.3 points above InternVL3-8B. Vad-R1 outperforms ARIA on seven of the eight metrics. This gap would be understandable because Vad-R1 uses SFT and reinforcement learning, whereas ARIA trains only lightweight routing components. Despite that, ARIA achieves a higher ROUGE-2 (0.206 vs. 0.194).

(2) Frame settings and ablation.

In the original configuration, Base samples 4 frames per segment, LOOK-MORE reruns the Base VLM with 8 frames on selected segments, and COOPERATE also uses 8 frames. We evaluate three Base/LOOK-MORE combinations (4/8, 4/16, and 8/16 frames) with COOPERATE excluded, both at a fixed routing threshold of τ = 0.20 and under a comparable budget of about 716k tokens per video.

Table 10. Ablation of Base/LOOK-MORE frame allocation.

Base / LOOK-MORE framesAUC (%) at τ = 0.20Tokens/videoAUC (%) at ≈716k budget
4/886.63 ± 0.07708k—
4/1686.82 ± 0.17765k85.60 ± 0.26
8/1687.34 ± 0.211,050k84.69 ± 0.03

At the fixed threshold, moving from 4/8 to 8/16 raises AUC by 0.71 points at about 48% higher token cost. When 4/16 and 8/16 are limited to about 716k tokens per video, both fall below 4/8. Using more frames per call therefore does not necessarily improve detection under a limited budget, and 4/8 gives a better balance between the visual input per call and the number of segments that can receive additional reasoning.

[1]Zou S, Tian X, Wesemann L, et al. Unlocking vision-language models for video anomaly detection via fine-grained prompting[C]//2026 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). IEEE, 2026: 4223-4233. [2]Huang C, Wang B, Wang W, et al. Vad-r1: Towards video anomaly reasoning via perception-to-cognition chain-of-thought[J]. Advances in neural information processing systems, 2026, 38: 118486-118518.

Comment 2: How does the number of input frames affect VAD performance?

Answer: We added a controlled experiment on how the number of input frames affects video anomaly detection. The Base VLM, prompts, UCF-Crime test set, and evaluation procedure are kept unchanged, and only the number of uniformly sampled frames per segment varies. ARIA's adaptive routing and additional reasoning actions are not used.

Table 11. Effect of the number of input frames on UCF-Crime.

FramesAUC (%) ↑Segment Acc. (%) ↑Latency ↓Visual tokens
281.2773.72.45512
483.7379.72.761,024
885.6482.23.562,048
1686.4480.15.234,096
3285.8179.29.348,192

Visual tokens are calculated at 256 tokens per frame. Latency is measured per segment in this fixed-frame experiment and does not directly represent the end-to-end latency of full ARIA.

AUC rises by 5.17 points from 2 to 16 frames and then drops at 32 frames, while segment accuracy peaks at 8 frames. Moving from 4 to 16 frames improves AUC by 2.71 points, but it quadruples the visual tokens and nearly doubles the latency. At 32 frames, the cost keeps growing without a further AUC gain.

Denser visual sampling can therefore improve VAD performance within a certain range, but these results do not support increasing the frame count indiscriminately for all segments. ARIA starts with fewer frames for low-cost initial inference, and then applies LOOK-MORE or COOPERATE to selected segments based on the expected correction benefit. We will report this result in the final version.

Comment 3: Can ARIA be evaluated on additional benchmarks?

Answer: Thank you for this suggestion. We added experiments on MSAD to further evaluate cross-dataset generalization. The router trained on the UCF-Crime training set is used directly, with the routing threshold kept at τ = 0.20 and no further training or MSAD-specific adaptation.

The Base VLM with 4-frame initial inference achieves 86.11% AUC on MSAD. ARIA achieves 91.49% and 91.38% in two runs, which is 5.38 and 5.27 points above the Base VLM. ARIA can therefore improve detection performance on a new dataset without retraining the router.

The revised manuscript will include the comparisons with ASK-HINT and Vad-R1, the frame configurations and both frame ablations with their computational budgets, and the MSAD experimental settings and results.

Dear Reviewer,

Thank you again for your valuable feedback. We have provided detailed responses and additional experimental results to address your concerns.

We would appreciate it if you could let us know whether our responses have addressed your main concerns, or if there are any remaining questions. We would be happy to provide further clarification during the discussion period.

Thank you for your time and consideration.

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

This paper presents ARIA, an adaptive reasoning framework for video anomaly detection. The main idea is simple and practical: instead of spending the same amount of computation on every video segment, ARIA decides whether to keep the initial prediction, look at more frames, or ask another VLM for help. I find the overall idea interesting, and the experimental results are encouraging, especially the improvement in inference efficiency. However, I still have a few concerns.

Critical strengths

Not all video segments need the same amount of reasoning. The idea of spending more computation only when needed makes sense and could be useful in real-world applications. ARIA can work with existing VLMs without fine-tuning them. The design is also easy to understand, and the two reasoning strategies, LOOK-MORE and COOPERATE, complement each other well.

Critical weaknesses
  1. The idea of using more computation for difficult samples is not entirely new. Also, looking at more frames and asking another model for help are both fairly straightforward strategies. I think the authors should better explain what makes ARIA different from existing adaptive inference and model routing methods.
  2. The Router is a key part of this paper. It is trained to predict whether additional reasoning can correct an initial mistake. However, in practice, it is also possible that the initial prediction is correct, but additional reasoning makes it wrong. Although the appendix provides some related analysis, I would like to see more details about how often the Router makes helpful or harmful decisions.
  3. To train the Router, the authors need to run all three inference strategies on the training data. This could be quite expensive. It would be helpful to report this extra cost and discuss whether the Router needs to be retrained when moving to a new dataset.
Questions for authors
  1. The paper reports good improvements in both accuracy and efficiency. While some alternative routing strategies are already compared in the appendix, I would like to see whether ARIA still performs better when different methods are given roughly the same inference budget. This would make the efficiency claim more convincing.
  2. The Router uses predictions from neighboring and visually similar video segments. This seems to require information from other parts of the video before making a decision. How would the method work for real-time video streams, where future frames are not yet available? I think this is worth discussing.
5 comments

Thank you for the valuable review. Our responses are addressed below. We use "Q" as an abbreviation for "Question".

Q1: Does ARIA outperform alternative routing strategies under the same inference budget?

Answer: Thank you for this suggestion. We added comparisons with simple routing strategies at the same token cost. The six baselines are Random, two confidence-based strategies, two temporal-change-based strategies, and Semantic-context entropy. All methods share the Base VLM, additional reasoning actions, action-selection head, and post-processing. Only the first-stage segment selection differs, and token consumption is kept the same across methods.

Table 6. Comparison of routing strategies at the same token cost on UCF-Crime.

Routing strategyRouted (%)AUC (%) ↑
Random38.286.88
Confidence: self-reported uncertainty36.386.70
Confidence: answer length42.185.82
Temporal: change from the previous segment38.786.72
Temporal: disagreement within ±2 segments38.986.58
Semantic-context entropy36.184.62
ARIA (ours)33.688.32

ARIA achieves the highest AUC while routing the fewest segments. Its router is trained on changes in prediction correctness before and after additional reasoning, whereas the baselines select segments at random or by heuristic signals and do not learn whether further reasoning can correct an initial error.

Q2: How can ARIA work in real-time streaming without future video information?

Answer: For the current version, ARIA is designed mainly for offline video anomaly detection, and it can be flexibly applied for real-time video stream setting. To specify, we tested a router whose context is built only from information available up to the current segment. Temporal Context uses the current and preceding segments, and Semantic Context retrieves only from segments already observed. The AUC of this past-only router is only 0.07% below the original ARIA in AUC, which suggests that future context is not necessary for the router to maintain good detection performance. Full streaming deployment would also require replacing centered sampling windows with historical windows and making semantic aggregation and temporal smoothing causal. In an additional experiment with causal post-processing, ARIA improves over the Base VLM by 6.60% in AUC under the same setting.

Table 7. Results with past-only router context and causal post-processing on UCF-Crime.

Experimental settingConfigurationAUC (%)Difference
Router contextOriginal ARIA (full context)88.32Reference
Router contextPast-only context88.25−0.07 pp
Causal post-processingBase VLM76.30Reference
Causal post-processingARIA82.90+6.60 pp

Comment 1: How often does ARIA's router make helpful versus harmful reasoning decisions?

Answer: Thank you for this suggestion. Additional reasoning can correct an initial mistake, but it can also turn a correct prediction into an incorrect one. To examine this, we compare ARIA with five simple segment-selection strategies. We define Wrong → Right as a helpful change and Right → Wrong as a harmful change. All percentages are computed over segments whose binary predictions change after additional reasoning.

Table 8. Helpful and harmful prediction changes under different routing strategies on UCF-Crime (%).

Routing strategyHelpful changes ↑Harmful changes ↓
Random43.456.6
Confidence: self-reported uncertainty47.352.7
Confidence: answer length47.852.2
Temporal: change from the previous segment49.150.9
Semantic-context entropy46.553.5
ARIA (ours)68.731.3

For all five simple strategies, fewer than half of the prediction changes are helpful, so they introduce more errors than they correct. For ARIA, 68.7% of the changes are helpful. These results suggest that learning from correction outcomes selects beneficial additional reasoning more effectively than random selection or heuristics based on self-reported uncertainty, temporal changes, answer length, or semantic disagreement. This analysis examines the outcomes of additional reasoning directly and provides evidence of the router's effectiveness beyond final AUC.

Comment 2: What is the Router's training cost, and does it require retraining on new datasets?

Answer: To clarify, the router itself and its training are light. According to our experimental records, this is a one-time offline cost of approximately 10 GPU-hours, and it is not repeated for each test video. At inference, Table 5 of the original paper reports an average of 5.9 minutes per video for ARIA, compared with 17.1 minutes for fixed 8-frame inference and 25.7 minutes for fixed cooperative inference. On retraining, the original paper reports that the router is trained only on UCF-Crime and applied directly to XD-Violence without further training or dataset-specific adaptation. We further tested its transferability on MSAD, using the same router with the threshold kept at τ = 0.20 and no retraining or adaptation. The AUC on MSAD is 91.49% and 91.38% across two random seeds, compared with 86.11% for the Base VLM. In the cross-dataset settings evaluated so far, the same trained router can therefore be applied to XD-Violence and MSAD without rebuilding the supervision data or retraining, which avoids repeated offline preparation costs.

We will add these results to the revised manuscript to address your constructive comment: the budget-controlled routing comparisons with the implementation of each baseline and the token accounting, the helpful and harmful change analysis with its statistical definitions, the past-only and causal experiments with a discussion of streaming deployment, and the offline cost and MSAD transfer results.

Dear Reviewer,

Thank you again for your valuable feedback. We have provided detailed responses and additional experimental results to address your concerns.

We would appreciate it if you could let us know whether our responses have addressed your main concerns, or if there are any remaining questions. We would be happy to provide further clarification during the discussion period.

Thank you for your time and consideration.

Reviewer_n19YOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ARIA, an adaptive reasoning framework for VLM-based video anomaly detection. A learned router dynamically selects among three actions: STOP, LOOK-MORE, and COOPERATE, based on initial predictions and contextual features. Experiments on UCF-Crime and XD-Violence demonstrate competitive detection performance with reduced token consumption and inference latency.

Critical strengths
  • This paper addresses the issue in VLM-based VAD, where different video segments often require varying degrees of visual evidence and reasoning. A unified solution is proposed.
  • The routing-and-action design adopted in the paper is intuitive and lightweight. Without the need to modify the VLM, an ideal balance between accuracy and efficiency is achieved.
Critical weaknesses
  • Adaptive computation, selective frame sampling, and token budgeting have been extensively explored in long/online video understanding. Related VAD works such as Holmes-VAU (CVPR 2025) and “Aligning Effective Tokens with Video Anomaly in Large Language Models”(ICCV 2025) are insufficiently discussed. ARIA primarily integrates existingtechniques rather than introducing a fundamentally new adaptive reasoning mechanism.
  • Although the VLM remains frozen, the training of the routing module relies on corrected labels generated based on frame-level ground truth. This supervisory requirement has not been adequately emphasized, and it may lead to an unfair comparison with baseline methods.
  • The experiments lack a comprehensive comparison with simpler adaptive strategies under the same token budget or delay constraints. Additionally, the routing module only prioritizes the corrected probability, without explicitly incorporating cost explanations for specific actions or considering the risk of converting correct predictions into incorrect ones.
Questions for authors
  • How does ARIA differ fundamentally from existing adaptive video sampling and token allocation methods?
  • How does ARIA perform when the router is trained using only video-level labels or pseudo-labels? How is supervision fairness ensured against training-free baselines?
  • Can the authors report action-specific correction rates, error introduction rates, and computational costs?
  • The paper reports improved explanation quality primarily through LLM-as-Judge evaluations. However, these metrics may not adequately capture visual grounding or hallucinations introduced by multi-model cooperation. Can the authors evaluate explanation faithfulness using benchmarks such as VALU (ACL 2026), or provide evidence-grounding analyses to verify whether COOPERATE improves factual accuracy rather than merely producing more plausible explanations?
7 comments

Thank you for the valuable review. Our responses are addressed below. We use "Q" as an abbreviation for "Question".

Q1: What distinguishes ARIA from existing adaptive video methods?

Answer: The fundamental difference between them and ARIA is that: unlike existing adaptive video sampling methods (e.g., Holmes-VAU) and token allocation methods (e.g., VA-GPT), which adaptively allocate computation before answer generation by selecting which video content to sample or retain, our ARIA operates after an initial prediction and focuses on whether additional reasoning can improve the answer. Rather than optimizing video content selection, ARIA identifies segments where the initial prediction may benefit from further reasoning and adaptively determines the appropriate reasoning action.

The key distinction is that Holmes-VAU and VA-GPT optimize information acquisition, whereas ARIA optimizes prediction refinement through targeted reasoning.

To specify, the contributions of ARIA are two-fold:

(1) Correctability-aware reasoning allocation. Unlike Holmes-VAU and VA-GPT, which allocate computation based on anomaly scores or visual feature changes, ARIA introduces a correctability-aware routing mechanism that learns whether additional reasoning can improve an initial prediction. By supervising the router with changes in prediction correctness before and after additional reasoning, ARIA allocates computation based on the potential for error correction rather than visual saliency.

In a controlled comparison where only the first-stage routing strategy differs, ARIA achieves 88.32% AUC while routing only 33.6% of segments, outperforming all heuristic baselines (best: 86.88% AUC with 38.2% routed segments). This demonstrates the effectiveness of learned correctability-aware routing over simple heuristic-based computation allocation.

(2) Lightweight training and cross-dataset generalization. Unlike Holmes-VAU and VA-GPT, which rely on supervised anomaly models and language-model fine-tuning or instruction tuning, ARIA keeps both VLMs frozen and trains only lightweight routing components (GBT heads and an auxiliary MLP).

More importantly, the router trained solely on UCF-Crime transfers directly to other datasets, including XD-Violence, without additional training or dataset-specific adaptation. This demonstrates that ARIA's reasoning allocation strategy can generalize across datasets without modifying the underlying VLMs.

Q2: How does ARIA perform under weak supervision, and is the comparison with training-free baselines fair?

Answer: Thank you for raising this important question. We clarify that ARIA keeps both VLMs frozen and trains only lightweight routing components. We further evaluate ARIA under weak supervision and clarify the supervision setting relative to training-free baselines.

(1) Performance with video-level labels and pseudo-labels. To assess ARIA's dependence on temporal annotations, we train the router using only video-level labels and segment-level pseudo-labels generated by majority voting over Base, LOOK-MORE, and COOPERATE predictions. No manually annotated anomaly intervals are used.

Table1. Performance of ARIA under different supervision settings on UCF-Crime.

Training SettingAUC (%)
Base VLM (No Router)83.73
Video-level Labels + Pseudo-label Router (3 seeds)87.15 ± 0.26
Full ARIA (Temporal Supervision)88.32

Without precise temporal supervision, ARIA achieves 87.15 ± 0.26% AUC, improving over the Base VLM by 3.42 percentage points and approaching the fully supervised ARIA (88.32%). This demonstrates that ARIA's routing mechanism remains effective with only video-level supervision and automatically generated pseudo-labels.

(2) Supervision fairness against training-free baselines. Video-level labels are standard annotations available in VAD benchmarks and are also utilized by existing approaches, such as VERA, for guiding-question construction. ARIA similarly leverages these available annotations, but uses them to train a lightweight router rather than fine-tuning the VLMs.

Thus, the key distinction lies in how the available supervision is utilized, rather than access to additional annotations. While ARIA is not strictly training-free at the routing level, its comparison with training-free baselines evaluates the effectiveness of lightweight reasoning allocation under standard VAD benchmark conditions, without requiring additional human annotations or backbone adaptation.

Q3: How does ARIA compare with simple routing methods, and what are the benefits, risks, and costs of its reasoning actions?

Answer: Thank you for this suggestion. We added comparisons with simple routing strategies under the same token budget, and measured the helpful and harmful prediction changes of each reasoning action together with their costs.

(1) Routing comparisons under the same token budget.

The results are in Table 2 as follows. Confidence-based routing ranks segments either by the frequency of uncertainty expressions (“maybe,” “possibly,” and “unclear”) or by answer length, with shorter answers treated as less confident. Temporal-change-based routing uses either the prediction change from the previous segment or the disagreement with the two neighboring segments on each side. Random and Semantic-context entropy are also included. All methods share the Base VLM, additional reasoning actions, action-selection head, and post-processing, so only the first-stage segment selection differs. ARIA exceeds every baseline by 1.44 to 3.70 percentage points in AUC. All baselines route more segments than ARIA but obtain lower AUC under the same token budget.

Table 2. Comparison of routing strategies under comparable inference budgets.

Routing strategyRouted segments (%)AUC (%)
Random38.286.88
Confidence: self-reported uncertainty36.386.70
Confidence: answer length42.185.82
Temporal: change from the previous segment38.786.72
Temporal: disagreement within ±2 segments38.986.58
Semantic-context entropy36.184.62
ARIA (ours)33.688.32

(2) Action-specific prediction changes and computational costs.

The percentages below are computed only over segments whose binary predictions change after additional reasoning.

Table 3. Action-specific correction and error rates.

MetricLOOK-MORECOOPERATEOverall
Helpful changes: Wrong → Right (%)75.866.068.7
Harmful changes: Right → Wrong (%)24.234.031.3

From Table 3, both actions produce more helpful corrections than harmful changes.

Table 4. Inference costs.

MetricValue
LOOK-MORE: additional tokens per selected segment2,576
COOPERATE: additional tokens per selected segment5,258
Full ARIA: average tokens per video716,512
Full ARIA: average end-to-end inference time per video5.9 min

LOOK-MORE is one additional Base VLM call with denser visual sampling. The full-system measurements use a batch size of 8, as in the original paper.

Q4: Can VALU verify whether ARIA improves explanation faithfulness rather than just plausibility?

Answer: Thank you for this suggestion. We added an evaluation on the Anomaly Localization and Description (ALD) task of VALU (ACL 2026), which evaluates temporal anomaly localization and descriptions at three semantic levels: Abnormal Event, Abnormal Segment, and Abnormal Action. We report VALU’s Coverage (Cov.) and Consistency (Con.) for all descriptions. Coverage measures how much of the reference anomaly information the generated description covers, and Consistency measures agreement between the generated facts and the human reference descriptions. These metrics assess factual consistency in more detail than the overall explanation scores in the original paper.

Table 5. Evaluation of anomaly localization and description faithfulness on VALU.

LevelMethod[email protected]mRmIoUCov.Con.
EventVideoLLaMA3-7B71.370.467.112.420.5
EventInternVL3-8B5.36.913.010.424.6
EventBase VLM (ours)13.015.218.67.522.0
EventARIA (ours)42.542.240.712.925.9
SegmentVideoLLaMA3-7B52.052.051.28.915.3
SegmentInternVL3-8B2.03.67.99.925.3
SegmentBase VLM (ours)18.322.629.27.116.4
SegmentARIA (ours)45.043.341.68.226.3
ActionVideoLLaMA3-7B14.017.824.98.912.3
ActionInternVL3-8B2.73.33.812.117.0
ActionBase VLM (ours)8.310.27.813.78.9
ActionARIA (ours)18.219.022.510.920.7

Our findings from this table are as follows:

(1) ARIA improves factual consistency at all levels.

From Table 5, ARIA’s Consistency is 25.9%, 26.3%, and 20.7% at the Event, Segment, and Action levels, the highest among all methods we compared at each level. The gains over our Base VLM are 3.9, 9.9, and 11.8 percentage points, and the gains over VideoLLaMA3-7B are 5.4, 11.0, and 8.4 points. ARIA’s improvements therefore extend beyond the original explanation quality scores to agreement with human reference anomaly facts.

(2) ARIA also improves temporal anomaly localization performance.

As shown in Table 5, ARIA’s mIoU is 40.7%, 41.6%, and 22.5% at the three levels, which is 22.1, 12.4, and 14.7 percentage points above the Base VLM. At the finer Action level, ARIA reaches 18.2% [email protected] and 19.0% mR, compared with 14.0% and 17.8% for VideoLLaMA3-7B.

Compared with the Base VLM without additional reasoning, ARIA improves both Consistency and mIoU at all three levels. These results support improvements in temporal anomaly localization and agreement with human reference facts, beyond producing plausible or fluent explanations.

In the revised manuscript, we will add the comparison with Holmes-VAU and VA-GPT, the routing comparisons under the same token budget with their implementation details and budget settings, the three training settings with their supervision and token consumption, the action-specific measurements with metric definitions, and the full VALU tables.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

Dear Reviewers and Area Chair,

Thank you for your time and detailed feedback on our paper. We have uploaded a revised version, with all new text highlighted in blue. Below, we summarize the main changes and indicate the reviewers who raised the corresponding points.

Main Text

  • Sec. 2 (Related Work): Added a paragraph comparing ARIA with Holmes-VAU and VA-GPT. These methods adapt visual input before answer generation, whereas ARIA decides after the initial prediction whether additional reasoning can correct an error. (n19Y)
  • Sec. 4.2 (Detection Performance): Added comparisons with ASK-HINT and Vad-R1. ARIA achieves 0.90 percentage points higher AUC than the reported ASK-HINT result on UCF-Crime, although the two methods use different model configurations. On Vad-Reasoning, ARIA outperforms all compared baselines except Vad-R1 in accuracy and F1. (AMjt)
  • Sec. 4.2 (Explanation Quality): Added an evaluation on VALU, which assesses temporal anomaly localization and description consistency against human references. ARIA improves both Consistency and mIoU over the Base VLM at the Event, Segment, and Action levels, providing evidence beyond LLM-based judging. (n19Y)
  • Sec. 4.3 (Router Training): Added a weakly supervised setting using video-level labels and automatically generated pseudo-labels. ARIA achieves 87.15 ± 0.26% AUC on UCF-Crime, compared with 83.73% for the Base VLM and 88.32% with temporal supervision. (n19Y)
  • Sec. 4.3 (Table 4): Added comparisons with six simple routing strategies under the same token budget. ARIA achieves the highest AUC (88.32%, compared with 86.88% for the best baseline) while routing the fewest segments. (n19Y, KkJA)
  • Sec. 4.4 (Performance–Cost Analysis): Added the average additional token cost per selected segment for LOOK-MORE (2,576) and COOPERATE (5,258), along with the one-time offline router training cost of approximately 10 GPU-hours. (n19Y, KkJA)
  • Sec. 4.5 (Generalization): Added cross-dataset evaluation on MSAD without retraining the router trained on UCF-Crime. ARIA achieves 91.49% and 91.38% AUC across two random seeds, compared with 86.11% for the Base VLM. (KkJA, AMjt)

Appendix

  • Secs. A.4.3–A.4.4: Added experiments on the effect of input frame count and an ablation of Base/LOOK-MORE frame allocation, including computational costs. (AMjt)
  • Sec. A.4.5: Added evaluations using past-only router context and causal post-processing to examine dependence on future information. (KkJA)
  • Sec. A.4.8: Added routing comparison details and analyses of helpful and harmful prediction changes after additional reasoning. (n19Y, KkJA)
  • Secs. A.4.11 and A.5.5: Added full experimental results on Vad-Reasoning and VALU, respectively. (AMjt, n19Y)

Further details are provided in our individual responses to each reviewer. We would appreciate any additional feedback and would be happy to clarify any remaining concerns during the discussion period.

Best,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.