acceptodds
Under review as a conference paper at ICLR 2027

DisCo: Distillation via Conflict for Weakly Supervised Video Anomaly Detection

Abstract

Video anomaly detection (VAD) has recently progressed along two promising directions: weakly supervised learning and adaptation of multimodal vision–language models (VLMs). Weakly supervised VAD is cost-effective because it relies only on video-level labels, but the lack of segment-level annotations makes it difficult to localize and disentangle multiple or diverse abnormal events. In contrast, VLMs pretrained on large-scale multimodal data encode rich, general semantic knowledge, yet they are not explicitly optimized for the domain-specific temporal reasoning required for anomaly detection. In this work, we propose a knowledge-sharing framework that transfers generalized semantics from VLMs to a weakly supervised VAD model via knowledge distillation. Empirically, we observe complementary behavior: zero-shot VLM predictions excel on certain anomaly categories (e.g., explosions and road accidents), whereas weakly supervised models are stronger on human-centric abnormalities (e.g., abuse). To enable selective distillation, we introduce a novel conflict score that quantifies class-wise disagreement by jointly modeling semantic and motion cues, and we distill knowledge only for classes with high conflict. Extensive experiments on two benchmark datasets show that our distillation via conflict (\bf DisCo) consistently improves anomaly detection performance over both weakly supervised and tuning-free VLM baselines, while requiring only a lightweight model at inference time.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Video anomaly detection, weakly supervised learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Xj2mOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3(+1)
Soundness
3(+1)
Read review
Paper summary

This paper introduces DisCo, a novel weakly supervised video anomaly detection (WSVAD) framework that employs knowledge distillation from a pretrained vision–language model (VLM). Based on observations from two widely used video anomaly detection benchmarks, UCF-Crime and XD-Violence, the authors identify complementary characteristics between motion-based anomaly detection and VLM predictions. They exploit this discrepancy to estimate the utility of VLM supervision and selectively distill semantic knowledge into a WSVAD model. Experimental results demonstrate substantial improvements over existing baselines in terms of Average Precision (AP) on both UCF-Crime and XD-Violence, providing empirical support for the effectiveness of the proposed selective distillation strategy. If the authors adequately address all the concerns raised in the Weaknesses section, I would be willing to increase my score.

Critical strengths

S1. Strong empirical performance with extensive ablation studies.

The proposed method demonstrates substantial improvements over existing baselines on both UCF-Crime and XD-Violence, particularly in terms of Average Precision (AP). The authors also conduct extensive ablation studies to investigate the effectiveness of individual components, including utility-based event selection, knowledge distillation strategies, and different teacher and student architectures. These experiments provide convincing evidence of the method's effectiveness on the evaluated benchmarks.

S2. Interesting motivation based on complementary characteristics of motion-based and semantic representations.

The observation that motion-based WSVAD models and VLMs exhibit complementary strengths in detecting different types of anomalies is interesting. The proposed strategy of selectively distilling semantic knowledge based on this discrepancy offers a potentially useful direction for improving WSVAD while retaining lightweight inference.

Critical weaknesses

W1. Insufficient evidence supporting the central motivation. (resolved)

The discrepancy between motion-based predictions and VLM confidence is central to this paper, yet the supporting evidence remains limited. Figures 1 and 4 mainly demonstrate this behavior on UCF-Crime, while Figure 2 shows category-wise performance differences rather than directly validating the proposed relationship.

I would raise my score if the authors could provide further quantitative evidence, particularly on XD-Violence, demonstrating that semantic-motion conflict reliably identifies beneficial distillation opportunities.

W2. Limited generalizability across teacher–student combinations. (resolved)

Although Appendices C.2.5 and C.2.7 provide some evidence of generalizability, I remain unconvinced that DisCo consistently works beyond the InternVL3-14B and UR-DMU combination. In particular, the improvements with alternative VLM teachers are considerably smaller.

I would consider raising my score if the authors could provide more systematic evidence across different teacher–student combinations.

W3. Dependence on event-level supervision and computational cost. (minor)

The framework relies on known event categories to estimate teacher utility and selectively distill knowledge. I question whether this approach remains effective under heavily imbalanced or evolving anomaly distributions, as commonly encountered in practice. Although Table 4 partially addresses this concern, the robustness of event-level selection remains unclear.

Additionally, offline VLM inference requires substantial computational resources, which may limit practical adaptability.

W4. Poor presentation and insufficient methodological justification. (resolved)

The introduction has significant presentation issues that substantially weaken the paper's overall impression. The authors repeatedly describe their methodological choices as "natural" without sufficiently explaining the underlying rationale. In particular, the motivation for combining VLMs with motion-based WSVAD models through knowledge distillation is not convincingly developed, despite being central to the proposed framework.

Furthermore, the distinction from previous work and the rationale for selecting UR-DMU as the primary student are insufficiently explained. Given the considerable unused space in the manuscript, I expected a more thorough introduction that clearly establishes the motivation, contributions, and methodological choices.

Although the experimental results are compelling, the weak presentation makes it difficult to fully appreciate the significance of the proposed approach. A substantially improved introduction would positively influence my assessment.

Questions for authors

See weaknesses.

4 comments

We would like to thank Reviewer Xj2m for reviewing and providing feedback on our work.

W1. Quantitative evidence supporting the central motivation:

We present analysis on test dataset in Table 10 to quantitatively validate our central motivation on both XD-Violence and UCF-Crime. We use the utility score (CrC_r) (Eq. 7) as the quantitative measurement for semantic-motion conflict (the same metric we use to select the event categories) for this analysis to demonstrate that it reliably identifies the gaps in the MIL-based learning that can be mitigated by VLM's semantic knowledge.

As discussed in the paper, the MIL-based learning overlooks subtle anomalies with low-motion changes, and this is where VLM's semantic knowledge is helpful. Table 10 reports recall for such low-motion segments i.e.segments whose motion change is below the median value. In the Table, UR-DMU represents MIL-based learning model, used as a base model in our paper.

Table 10. Recall on low-motion anomalies across semantic–motion utility groups on XD-Violence and UCF-Crime. Bold values indicate the better performance between the VLM and UR-DMU. Values in parentheses indicate gains over UR-DMU.

DatasetUtility ScoreVLMUR-DMUDisCo
XD-ViolenceTop 20%89.476.190.5 (+14.4)
XD-ViolenceBottom 20%35.048.255.5 (+7.3)
XD-ViolenceOverall70.164.175.4 (+11.3)
UCF-CrimeTop 20%84.639.750.0 (+10.3)
UCF-CrimeBottom 20%3.81.625.8 (+24.2)
UCF-CrimeOverall33.215.729.9 (+14.2)

(1) MIL-based learning struggles with low-motion anomalies. UR-DMU achieves only 15.7% and 64.1% recall on low-motion anomalies in UCF-Crime and XD-Violence, respectively, highlighting the difficulty of detecting subtle anomalies using motion-dominated MIL supervision.

(2) VLMs provide complementary semantic evidence. On these low-motion anomalies, the VLM achieves higher recall than UR-DMU on both datasets: 33.2% vs. 15.7% (+17.5 points) on UCF-Crime and 70.1% vs. 64.1% (+6.0 points) on XD-Violence. These results quantitatively support our motivation that VLMs can recognize semantically identifiable anomalies that motion-based learning may overlook.

(3) Semantic–motion utility identifies useful VLM evidence.

On XD-Violence, for videos with the top 20% utility scores, the VLM's recall is higher (89.4%) than UR-DMU (76.1%), indicating that transferring knowledge from the VLM could be beneficial. However, for videos with the bottom 20% utility scores, the VLM's recall is lower (35.0%) than UR-DMU (48.2%), suggesting that knowledge transfer may be less beneficial in such cases. Similarly, on UCF-Crime, the VLM's recall is substantially higher (84.6%) than UR-DMU (39.7%) for the top 20% utility group, while for the bottom 20%, its recall is very low (3.8%), indicating that the VLM provides limited reliable supervision in such cases. This indicates that higher utility scores are associated with more informative VLM supervision, supporting their use for identifying beneficial distillation opportunities.

(4) DisCo improves low-motion anomaly detection. DisCo increases overall low-motion anomaly recall over UR-DMU by 11.3 percentage points on XD-Violence (64.1% to 75.4%) and 14.2 points on UCF-Crime (15.7% to 29.9%). Importantly, in the top-utility groups, DisCo improves recall by 14.4 points on XD-Violence and 10.3 points on UCF-Crime, demonstrating improved detection in regions where the VLM provides complementary semantic evidence.

Interestingly, DisCo also improves recall in low-utility regions, particularly on UCF-Crime (+24.2 points). Our category-level analysis indicates that this improvement arises predominantly from Arson videos, despite Arson being excluded from direct event-wise distillation due to a lower utility score. One possible explanation is that distillation on selected categories such as Explosion enables the student to generalize shared visual patterns, such as fire and smoke, to related Arson.

Moreover, we have also revised Figure 4 to include the relationship between anomaly score and motion on XD-Violence.

W2. Limited generalizability across teacher–student combinations:

Please refer to our response to Reviewer fBk7, W3.

We conducted experiments with different teacher models using event-specific prompts (our final configuration) and also evaluated vanilla KD with these teachers. Moreover, our method generalizes to feature families beyond I3D, such as CLIP, and also provides additional gain from other distillation-based methods (DAKD and π\pi-VAD) as reported in the table below.

Table 11. Compatibility of DisCo with different top-kk MIL-based weakly supervised methods. Encoder denotes the student inference encoder. Performance is reported in AP (%).

MethodEncoderAdditional SupervisionUCF-Crime OriginalUCF-Crime + DisCoUCF-Crime GainXD-Violence OriginalXD-Violence + DisCoXD-Violence Gain
UR-DMUI3DNone35.4842.29+6.8179.1485.96+6.82
VadCLIPCLIPText Semantics33.5540.15+6.6084.5086.07+1.57
PE-MILI3DSemantic Prompt Learning33.4138.81+5.4087.2789.06+1.79
π\pi-VADI3DMulti-encoder KD33.9938.54+4.5582.6586.64+3.99
DAKDI3DMulti-encoder KD29.4137.36+7.9583.7186.83+3.12

W3. Dependence on event-level supervision and computational cost

Both UCF-Crime and XD-Violence have unequal video counts across event categories. To assess selection robustness, we vary event-level imbalance on UCF-Crime from 1:1 to 5.1:1 (largest-to-smallest category ratio). Table 12 reports results using fixed utility-estimation budgets of 351 (27 per category, matching the smallest original category) and 500 videos. We recompute event selection under each distribution while keeping student training unchanged.

Table 12. Robustness of DisCo under varying event-class imbalance on UCF-Crime. OR denotes overlap with the event selected by DisCo on full training videos.

BudgetDistributionImbalanceOR (%)AP (%)
351Balanced1:180.838.87±0.3538.87 \pm 0.35
351Mild2.0:179.639.74±1.3639.74 \pm 1.36
351Moderate3.0:178.139.30±0.7439.30 \pm 0.74
351Severe5.1:173.938.65±0.9338.65 \pm 0.93
500Balanced1:184.740.98±1.8040.98 \pm 1.80
500Mild2.0:186.138.97±0.5238.97 \pm 0.52
500Moderate3.0:182.539.14±1.8939.14 \pm 1.89
500Severe5.1:178.237.77±1.0737.77 \pm 1.07
DisCo (full data, 810 videos)10042.29
Vanilla KD–35.94
Baseline–35.48

Across different imbalance levels, DisCo maintains over 70% selection overlap with the full-data reference, retaining approximately 4–5 of the six reference events on average. Although reducing the utility-estimation data affects selection stability and performance, DisCo consistently outperforms vanilla KD (35.94% AP) across all evaluated settings. Under the strongest imbalance (5.1:1), where some categories contain as few as 12 videos for utility estimation, DisCo achieves 38.65% and 37.77% AP with 351 and 500 videos, respectively, exceeding vanilla KD by 2.71 and 1.83 points.

Our analysis considers changes in the frequency of known event categories. For settings without event labels or with evolving anomaly types, DisCo can instead group videos by clustering VLM-generated segment descriptions. This fully event-agnostic variant (Table 4 of the main paper) still achieves 39.27 AP, compared with 35.48 for the baseline and 35.94 for vanilla KD.

Computational cost: The VLM inference cost is a one-time offline pre-processing step, with teacher predictions computed and cached before student training. Once pre-processed, the VLM is not required during each student training iteration. Crucially, after training the student model, DisCo requires no additional VLM inference overhead at test time or in its regular deployment. Thus, the computational expense is confined to offline knowledge extraction, while deployment retains the efficiency of the original lightweight student model.

The total GPU-hours required for different teacher models are reported below.

Table 13. Approximate offline VLM inference cost for five runs on UCF-Crime.

TeacherGPU-hours (5 runs)Seconds/ segmentSeconds/ video
InternVL3-14B540.5848
InternVL3-9B24.60.2721.8
LLaVA-NeXT-Video-7B20.60.2218.3
Qwen3-VL-8B17.40.1915.5

W4 Presentation: We have substantially revised the introduction (highlighted in blue in the revised manuscript, including Figure 2), particularly the motivation for complementary VLM–WSVAD learning and the rationale for selective knowledge distillation, following the analysis presented in our response to W1.

Reviewer_Xj2mOpenReview ↗

Thank You for the Thorough Rebuttal!

Thank you for the comprehensive rebuttal and substantial revisions.

I have carefully reviewed the additional experiments and revised manuscript. All of my original concerns (W1–W4) have been satisfactorily resolved. I appreciate the authors' efforts in strengthening the paper.

I am therefore increasing my rating from 2 to 3 (Weak Accept) and remain open to further upward revision during the discussion period.

Good job, and good luck with the rest of your research!

AuthorsOpenReview ↗

Thank you!

Thank you for your valuable feedback and for taking the time to carefully review our rebuttal. The additional experiments and analyses you suggested have helped strengthen our paper. We greatly appreciate your positive reassessment of our work and decision to increase the rating. Thank you again for your support.

Reviewer_fBk7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to transfer semantic knowledge from a frozen vision-language model to a weakly supervised video anomaly detector while retaining only the student model at inference time. DisCo computes regional utility scores from the conflict between the teacher's anomaly predictions and changes in visual features. It then aggregates these scores by event category and selects suitable anomaly categories for teacher-to-student distillation. Using InternVL3-14B as the teacher and UR-DMU as the student, the method achieves AP scores of 42.29 and 85.96 on UCF-Crime and XD-Violence, respectively. The corresponding student baselines achieve 35.48 and 79.14, while vanilla KD achieves 35.94 and 82.50. Experiments examine selection criteria, grouping levels, dependence on event labels, and teacher and student variants to analyze the role of selective knowledge transfer.

Critical strengths

1.Selective distillation has substantive experimental support. Compared with vanilla KD, DisCo improves AP by 6.35 and 3.46 percentage points on the two datasets, respectively. In Table 3, event-level and video-level selection use the same number of videos, 338, and achieve UCF AP scores of 42.29 and 34.48, respectively. This comparison provides further support for event-level aggregation. 2.The method has practical deployment value and examines dependence on event labels. Only the student is retained at inference time, avoiding the need to run a large model online. The variant without event labels still achieves 39.27 AP, compared with 35.48 for the student baseline, indicating that some gains can be obtained through automatic semantic grouping. This experiment helps distinguish the contribution of the selection mechanism from that of manually provided event-category priors.

Critical weaknesses

1.The incremental contribution over closely related work is not sufficiently distinguished. DAKD has already studied transferring knowledge from a multi-backbone teacher to a single-backbone anomaly detector. π-VAD has injected multimodal knowledge into UR-DMU while using RGB inputs at inference time. The paper does not discuss the former and cites the latter without a numerical comparison. DisCo's incremental contribution should center on the conflict between semantic predictions and feature changes, event-level aggregation, and selective distillation. The paper includes selection and aggregation ablations, but a comparison against simple teacher-score ranking is still needed, with the number of selected categories fixed and the amount of distillation supervision matched as closely as possible. This would clarify the additional value of the conflict mechanism. The distinction from closely related methods and the necessity of the core mechanism remain insufficiently established, limiting the assessment of novelty. 2.Coverage of strong baselines is incomplete, and competitiveness has not been sufficiently established. The main comparison omits relevant, strong published methods. The RGB/I3D version of PE-MIL and the I3D version of LEC-VAD report XD AP scores of 88.05 and 88.47, respectively. The full version of DSANet achieves 86.95, whereas this paper reports only its restricted configuration at 82.88, with the restriction disclosed. DisCo achieves 85.96. These published results come from different experimental pipelines. They should be included alongside explanations of differences in features, supervision, post-processing, and inference components so that competitiveness can be assessed. Results reported in the original papers should be clearly distinguished from reproductions under a common protocol. Priority should be given to full baseline configurations directly relevant to this paper's setting. 3.Evidence for cross-model generalization and the stability of the gains remains limited. The appendix evaluates three teachers and two students, but the students are primarily based on I3D features. These experiments use definition-guided prompts rather than the event-specific prompts used for the main results. The smaller teachers yield UCF AP scores of 35.75 and 35.89, close to the student baseline of 35.48. Without independent repeated training runs, it is difficult to determine whether these gains exceed training variability. The current results provide only preliminary support for effectiveness across a broader range of teachers and students. I recommend adding a student from a different feature family under the final configuration, aligning the settings of the existing teacher variants, and conducting paired repeated-training experiments for the student baseline, vanilla KD, and DisCo. This evidence would directly affect my assessment of the method's applicability and the reliability of its contribution.

Questions for authors

1.Contribution and mechanism: What are the additional mechanisms and conditions of applicability relative to DAKD and π-VAD? With the number of selected categories fixed and the amount of distillation supervision matched as closely as possible, does conflict-based ranking outperform ranking solely by the teacher's mean anomaly probability? 2.Completeness of comparisons: Could the authors add PE-MIL, LEC-VAD, and the full DSANet, distinguish originally reported results from local reproductions, and explain differences in features, supervision, and inference components? 3.Generalization and stability: Under the final prompt configuration, does a student from a different feature family retain the gains over vanilla KD? Please report paired independent repeated-training results and clarify whether the gains from the smaller teachers exceed training variability. The priority is representative configurations, rather than an exhaustive grid of teacher-student combinations. Training and model selection: What training or validation information is used to select hyperparameters and checkpoints? Is the MIL weight for unselected videos 1 or 0.1? Do vanilla KD and DisCo use the same auxiliary losses, effective loss weights, and checkpoint selection rules? 4.Metric trade-offs and implementation: UCF AP improves, but AUROC decreases from 86.97 to 86.44. Please explain this trade-off. R@1/2/3 denotes recall at FPRs of 1%/2%/3%. Since five binary teacher votes produce only six distinct raw scores, is recall computed as the maximum achievable recall under the FPR constraint, from the nearest ROC point, or by interpolation? Please provide the curves and calculation rules.

4 comments

We would like to thank Reviewer fBk7 for the detailed feedback on our work. We believe the points raised are helpful to strengthen our paper.

W1. Contribution and mechanism:

W1.1 Novelty over baselines carrying out distillation from multiple backbones

DAKD and π\pi-VAD primarily focus on aggregating representations from multiple pretrained encoders and distilling them into a student operating on a single encoder. However, their representations are still optimized using MIL-based objectives (top-1 ranking in DAKD and top-kk MIL in π\pi-VAD), making them susceptible to the limitations discussed in Lines 79–83, where supervision favors dominant anomalies while overlooking subtle anomalies. In contrast, DisCo transfers complementary semantic knowledge from a general-purpose VLM pretrained on large-scale, diverse multimodal data, providing supervision beyond MIL-based representations. Crucially, our conflict-aware mechanism identifies where semantic knowledge complements motion-based learning, and event-level selection enables targeted distillation while avoiding unreliable teacher predictions. This distinction constitutes the central contribution of DisCo over conventional representation distillation.

The effectiveness of distilling from a general-purpose VLM is further demonstrated by applying DisCo to DAKD and π\pi-VAD. As shown in Table 3, DisCo improves DAKD’s AP by 7.95 and 3.12 points on UCF-Crime and XD-Violence, respectively, while improving π\pi-VAD's AP by 4.55 and 3.99 points on UCF-Crime and XD-Violence, respectively. These consistent improvements demonstrate that DisCo provides complementary semantic supervision even to models already benefiting from multi-encoder knowledge distillation, highlighting its generalizability across different WSVAD frameworks. Since the official code for DAKD, π\pi-VAD, and PE-MIL is not publicly available, the reported results are from our reproduction. We note that the original papers report 85.12 AP (DAKD) and 85.37 AP (π\pi-VAD) on XD-Violence, whereas our reproductions obtain 83.71 and 82.65, respectively. DisCo surpasses these originally reported results as well.

Table 3. Comparison with knowledge distillation-based methods on UCF-Crime and XD-Violence.

ModelEncoderUCF-Crime (%) APUCF-Crime (%) R@1UCF-Crime (%) R@2UCF-Crime (%) R@3XD-Violence (%) APXD-Violence (%) R@1XD-Violence (%) R@2XD-Violence (%) R@3
DAKDI3D29.4112.6918.0823.5283.7138.0448.1154.29
DAKD + DisCoI3D37.3615.1423.5130.3986.8345.2757.4664.78
π\pi-VADI3D33.9914.8321.5626.1582.6534.9946.4653.87
π\pi-VAD + DisCoI3D38.5415.9525.0632.5486.6443.0156.7564.35
UR-DMUI3D35.4817.0017.0021.2079.1426.9338.6746.56
UR-DMU + DisCoI3D42.2922.8730.3234.5785.9644.9955.3362.22

W1.2 Ranking by teacher's mean anomaly prediction: To isolate the contribution of semantic–motion conflict, we compare our selection criterion against a simpler baseline that ranks events by the teacher's mean anomaly prediction, while keeping the number of selected categories fixed at five and using the same distillation objective. As shown in the table below, selection based on the teacher's mean anomaly score achieves 80.62 AP on XD-Violence, whereas our conflict-based selection achieves 85.96 AP, an improvement of 5.34 percentage points. Notably, teacher-score ranking excludes Abuse and Car Accidents, while our method excludes Abuse and Riot. These results demonstrate that high teacher confidence alone is insufficient to identify where distillation is most beneficial; explicitly accounting for semantic–motion disagreement provides a more effective criterion for selecting events where the VLM offers complementary supervision.

Table 4. Comparison of event selection based on teacher ranking on XD-Violence

CriterionAPR@1R@2R@3Event Selection Remarks
Baseline79.1426.9338.6746.56No distillation
Teacher's mean anomaly score80.6228.8242.6251.65Excludes Abuse and Car Accidents
Semantic–motion disagreement85.9644.9955.3362.22Excludes Abuse and Riot

W2. Completeness of comparisons: We thank the reviewer for highlighting these additional baselines and have included the reported results of LEC-VAD and the full DSANet, together with our reproduction of PE-MIL (the original paper reports 88.05 AP on XD-Violence; our reproduction obtains 87.27). Regarding PE-MIL, we acknowledge that it incorporates complementary semantic knowledge through prompt learning. However, its prompt learning relies on anomaly scores to distinguish potential anomalous events from surrounding normal segments within abnormal videos. Since these scores are learned through MIL-based supervision, they may inherit the underlying bias toward high-scoring segments, potentially overlooking subtle anomalies. Our results further support this limitation: incorporating DisCo's selective semantic distillation into PE-MIL improves AP from 33.41 to 38.81 (+5.40) on UCF-Crime and from 87.27 to 89.06 (+1.79) on XD-Violence, surpassing the reported results of both LEC-VAD (88.47) and the full DSANet (86.95) on XD-Violence. These improvements demonstrate that DisCo provides complementary semantic supervision even to methods that already incorporate semantic knowledge, highlighting its effectiveness and applicability beyond the UR-DMU backbone.

Table 5. Additional baselines on UCF-Crime and XD-Violence.

ModelEncoderUCF-Crime (%) APUCF-Crime (%) R@1UCF-Crime (%) R@2UCF-Crime (%) R@3XD-Violence (%) APXD-Violence (%) R@1XD-Violence (%) R@2XD-Violence (%) R@3
LEC-VADI3D+Text––––88.47–––
DSANetViT37.41–––86.95–––
PE-MILI3D+Text33.4113.0519.5825.987.2747.456.5163.97
PE-MIL + DisCoI3D+Text38.8117.0326.7733.5789.0653.2962.7268.53

W3. Cross-teacher and cross-student generalization To provide stronger evidence of generalizability, we extend our experiments to four VLM teachers and two student architectures from distinct feature families (I3D and CLIP), using consistent event-specific prompting and three independent training runs. As shown in Table 6, DisCo consistently outperforms vanilla KD across all eight teacher–student combinations, with AP improvements ranging from 3.88 to 9.55 percentage points. Notably, with smaller VLM teachers, indiscriminate distillation can harm the student, reducing performance below the baseline. In contrast, DisCo remains above the corresponding baseline in all eight settings. This shows that conflict-aware selection prevents unreliable teacher knowledge from degrading the student, even when the teacher is weaker. Furthermore, the reported mean and standard deviation across repeated runs show that DisCo's improvements over vanilla KD are consistently positive, providing stronger evidence of the reliability and generalizability of our selective distillation framework.

Table 6. Cross-teacher generalization of DisCo on UCF-Crime.

StudentEncoderTeacherBaselineVanilla KDDisCo
UR-DMUI3DInternVL3-14B33.71±2.1533.71 \pm 2.1535.60±0.7035.60 \pm 0.7040.33±1.92\mathbf{40.33} \pm 1.92
UR-DMUI3DInternVL3-9B33.71±2.1533.71 \pm 2.1530.36±0.5930.36 \pm 0.5937.54±1.53\mathbf{37.54} \pm 1.53
UR-DMUI3DQwen3-VL-8B33.71±2.1533.71 \pm 2.1531.22±0.4731.22 \pm 0.4735.10±0.86\mathbf{35.10} \pm 0.86
UR-DMUI3DLLaVA-NeXT-7B33.71±2.1533.71 \pm 2.1529.45±0.8529.45 \pm 0.8536.41±0.94\mathbf{36.41} \pm 0.94
VadCLIPCLIPInternVL3-14B28.51±2.9828.51 \pm 2.9833.14±2.4933.14 \pm 2.4937.99±1.93\mathbf{37.99} \pm 1.93
VadCLIPCLIPInternVL3-9B28.51±2.9828.51 \pm 2.9822.12±0.4822.12 \pm 0.4831.67±1.12\mathbf{31.67} \pm 1.12
VadCLIPCLIPQwen3-VL-8B28.51±2.9828.51 \pm 2.9828.62±1.7928.62 \pm 1.7933.00±2.81\mathbf{33.00} \pm 2.81
VadCLIPCLIPLLaVA-NeXT-7B28.51±2.9828.51 \pm 2.9828.20±3.4228.20 \pm 3.4233.67±3.51\mathbf{33.67}\pm 3.51

Q3. Training and model selection:

In WSVAD benchmarks, training videos lack frame-level annotations, therefore we follow common practice and report results on the standard test set. To further assess the effect of checkpoint selection, we conducted an additional experiment on UCF-Crime, where we randomly reserved 20% of the test videos as a validation set for checkpoint selection and evaluated the selected checkpoint on the remaining 80% of test videos as well as the full testset.

Table 7. AP comparison on UCF-Crime using validation-based checkpoint selection across five random 20% validation splits.

MethodDisjoint Test AP (%)Full Test AP (%)
UR-DMU26.95±3.0926.95 \pm 3.0931.41±0.8631.41 \pm 0.86
Vanilla KD25.55±6.7125.55 \pm 6.7132.08±4.0732.08 \pm 4.07
DisCo (Ours)35.57±1.86\mathbf{35.57} \pm 1.8640.82±1.81\mathbf{40.82} \pm 1.81

As shown in Table 7, DisCo achieves 35.57±1.8635.57\pm1.86 AP on the disjoint evaluation subset, outperforming UR-DMU and vanilla KD by 8.62 and 10.02 AP points, respectively. The improvement is also maintained when evaluating the validation-selected checkpoints on the full test set.

The hyperparameter β=0.3\beta=0.3, used in our main experiments, also achieves the highest validation and disjoint-test AP under validation-based selection (Table 8).

Vanilla KD and DisCo share the same auxiliary losses and checkpoint rule.

Table 8. Sensitivity analysis of β\beta on UCF-Crime using validation-based selection across five random 20% validation splits.

β\boldsymbol{\beta}Validation AP (%)Disjoint Test AP (%)Full Test AP (%)
0.055.28±6.5255.28 \pm 6.5231.91±5.3231.91 \pm 5.3237.36±2.4937.36 \pm 2.49
0.357.23±6.14\mathbf{57.23} \pm 6.1435.57±1.86\mathbf{35.57} \pm 1.8640.82±1.81\mathbf{40.82} \pm 1.81
0.553.58±6.9853.58 \pm 6.9831.88±3.7331.88 \pm 3.7336.66±1.1136.66 \pm 1.11
0.753.11±7.1453.11 \pm 7.1434.10±1.5434.10 \pm 1.5438.18±1.3038.18 \pm 1.30
1.048.73±5.0848.73 \pm 5.0832.91±2.6032.91 \pm 2.6036.23±1.1536.23 \pm 1.15

MIL Weighting and Consistency of Training Objectives

We have revised the Sec. 3.5 Training Objective section and Appendix B.1 to explicitly specify the event-dependent MIL and distillation weights, their normalization, and the objective for vanilla KD.

The event-dependent parameter λe\lambda_e determines the MIL and distillation weights as (1−λe)(1-\lambda_e) and λe\lambda_e, respectively. Selected events use (λe=λs)=0.9(\lambda_e=\lambda_s)=0.9, while unselected events use (λe=λu)=0(\lambda_e=\lambda_u)=0, retaining MIL supervision without distillation. So the MIL weight for unselected videos is 1.

Q4. Metric trade-offs:

AUROC performance

We clarify that the previously reported UR-DMU AUROC of 86.97% was taken from the original paper rather than our reproduction.

To ensure a fair comparison, we evaluated both UR-DMU and DisCo across three random seeds under the same experimental setting. DisCo improves AP from 33.71±2.1533.71\pm2.15 to 40.33±1.9240.33\pm1.92 (+6.62 points), while maintaining comparable AUROC as shown in Table 9.

Table 9. AP and AUROC comparison on UCF-Crime.

MethodAP (%)AUROC (%)
UR-DMU33.71±2.1533.71 \pm 2.1585.76±0.7085.76\pm 0.70
DisCo (Ours)40.33±1.9240.33 \pm 1.9286.15±0.3086.15 \pm 0.30

R@k\boldsymbol{R@k} denotes frame-level recall at a false positive rate (FPR) of k%k\%. We sort the anomaly scores of normal test frames in descending order and set the threshold to the ⌈(k/100)N⌉\lceil (k/100)N \rceil-th highest score, where NN is the number of normal frames. Every frame with a score greater than or equal to this threshold is classified as anomalous, and recall is computed at that threshold. Equivalently, R@kk corresponds to the first ROC point whose FPR reaches k%k\%, without interpolation.

We have revised the main result table and Sections C.2.1, C.2.5, C.2.7, and C.3.7 of the paper to incorporate the rebuttal experiments and additional analyses.

We sincerely thank you for your thoughtful feedback and constructive suggestions. We have carefully addressed your concerns regarding novelty, additional baseline comparisons, generalizability, and methodological clarity through further experiments and detailed clarifications. In particular, for generalizability, you can also refer to Table 11 in our response to Reviewer Xj2m, where we distinguish the different forms of supervision used by student baselines and highlight the consistent performance improvements achieved by DisCo across these baselines.

We hope these additional experiments and clarifications adequately address your concerns. If you have any remaining questions or concerns, we would greatly appreciate the opportunity to address them. We would also be grateful if you would consider revisiting your assessment in light of these additional results.

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

The paper proposes DisCo, a selective distillation framework for weakly supervised video anomaly detection. It addresses the uneven usefulness of VLM supervision across anomaly categories. DisCo estimates regional teacher utility from semantic–motion conflict, aggregates utility within event categories, and distills teacher predictions only for selected categories. The student operates independently at inference. Experiments on UCF-Crime and XD-Violence show substantial AP improvements over the student baseline and indiscriminate distillation. Comparisons under matched prompts, video counts, and category counts support the contribution of the proposed supervision selection strategy.

Critical strengths
  1. The paper identifies a consequential distinction between teacher capability and supervision utility. A VLM can provide valuable knowledge for particular events even when its overall anomaly-detection performance is weaker than the student’s. DisCo makes this uneven complementarity actionable through selective supervision. This advances the use of imperfect teachers by addressing where their knowledge can improve the learner, a question that overall teacher accuracy alone cannot answer.
  2. The paper develops a meaningful approach to choosing the unit of supervision transfer. Event-level selection pools evidence across semantically related videos to make transfer decisions that would otherwise depend on noisy local disagreement. Together with the semantic–motion criterion, this provides a concrete approach to selecting supervision under weak labels. The targeted comparisons support this methodological contribution and strengthen the case that the work offers more than an application of standard distillation.
Critical weaknesses
  1. The utility score lacks a sufficiently clear basis for distinguishing complementary semantic knowledge from confident teacher errors. Event aggregation can stabilize the score while retaining systematic errors, limiting the justification for treating conflict as transfer utility.
  2. The effective training objectives across selective and indiscriminate distillation are insufficiently specified. This makes it difficult to separate the benefit of selecting better supervision from changes in the relative influence of the learning objectives.
Questions for authors
  1. Please clarify how the utility score distinguishes useful semantic correction from confident teacher errors, and how event aggregation affects this distinction.
  2. Please specify the effective training objectives for selected videos, unselected videos, and indiscriminate distillation, including MIL weighting and distillation normalization.
3 comments

We thank Reviewer ZuYL for raising these important concerns and acknowledging the contributions of our work.

W1. Distinguishing Semantic Corrections from Teacher Errors through Utility Scoring and Event Aggregation

Our utility score is designed to prioritize the recovery of potentially missed abnormal segments by MIL-based training, as missed anomalies (false negatives) incur comparatively higher security cost than having false alarms. We distinguish two possible effects of VLM predictions on ground-truth abnormal segments: (i) beneficial semantic correction, where the VLM identifies anomalies in low-motion regions that may be underexplored by MIL-based learning; and (ii) harmful teacher errors, where the VLM confidently makes error for predicting abnormal segments, potentially reinforcing false negatives. Since segment-level ground truth is unavailable during training, our utility score does not explicitly identify correct and incorrect teacher predictions. Instead, it uses the direction and magnitude of semantic–motion disagreement to prioritize potential corrections while reducing the influence of certain harmful predictions. To explain this behavior, we conduct a post-hoc diagnostic analysis (Table 1) of ground-truth abnormal regions from events selected and excluded by DisCo.

Prioritizing complementary semantic evidence.

When motion evidence is weak but the VLM confidently predicts the region as abnormal (prteach=1\boldsymbol{p^{\mathrm{teach}}_r = 1}), the resulting semantic-dominant disagreement receives higher utility (Eq. 5). For example, in the selected event Explosion, 26.7% of ground-truth abnormal regions exhibit low motion but receive confident, correct VLM predictions, demonstrating complementary semantic evidence where relative motion cues are weak.

Limiting harmful teacher errors.

The utility score reduces the influence of confidently incorrect VLM predictions (prteach=0\boldsymbol{p^{\mathrm{teach}}_r = 0} on abnormal regions) through two mechanisms. (1) Low-motion regions: When the VLM confidently predicts a low-motion abnormal region as normal, both motion and semantic signals indicate weak anomaly evidence. Consequently, their disagreement is small (Eq. 4), resulting in a low utility score. In Arson, 41.5% of ground-truth abnormal regions exhibit this pattern, whereas only 1.2% exhibit correct VLM predictions in low-motion regions. Thus, despite recognizing some higher-motion abnormalities (12.9%), the teacher provides limited complementary evidence in low-motion regions, reducing the utility of transferring its predictions. (2) Higher-motion regions: When the VLM confidently predicts a region to be normal despite stronger motion evidence, the resulting motion-dominant disagreement is explicitly downweighted by the asymmetric factor (β\beta, Eq. 7). In Shoplifting, 25.6% of abnormal regions exhibit this pattern, while none (0.0%) of the low-motion abnormal regions receive correct VLM predictions. Thus, transfer of confident errors in Arson and Shoplifting is mitigated.

Table 1. VLM prediction behavior on ground-truth abnormal regions across different motion conditions. Percentages are relative to all abnormal segments in the event category.

EventMotion# Abnormalprteach=1\boldsymbol{p^{\mathrm{teach}}_r = 1} ↑\uparrowprteach=0\boldsymbol{p^{\mathrm{teach}}_r = 0} ↓\downarrowRemarks
ArsonLow1021.2%41.5%Unselected Event
ArsonHigh6912.9%5.3%
ExplosionLow8626.7%11.8%Selected Event
ExplosionHigh7528.6%2.5%
ShopliftingLow290.0%5.0%Unselected Event
ShopliftingHigh1311.3%25.6%

The effectiveness of the utility score in mitigating confident teacher errors while prioritizing complementary semantic evidence is further supported by the selection-criterion ablation (Sec. C.2.1). Teacher-reliability-based selection ranks Shoplifting among the top three events despite its frequent confident false negatives, achieving 35.19% AP, compared with 42.29% AP using our utility-based selection.

Event-level aggregation stabilizes the utility estimate but does not, by itself, eliminate teacher errors. Since teacher errors exhibit event-dependent patterns (Fig. 2), recurring errors that receive low regional utility can reduce the event-level mean utility, allowing DisCo to exclude the affected events from distillation. To examine this effect, we extend our post-hoc analysis to compare VLM predictions across events selected and unselected by DisCo (Table 2). In unselected events, the VLM confidently misses 56.2% of low-motion abnormal regions and correctly identifies only 6.7%. In contrast, for selected events, the VLM correctly identifies 32.7% of low-motion abnormal regions, supporting the decision to distill complementary semantic knowledge from these events. Selected events also exhibit fewer confident teacher errors in both low-motion (25.9% vs. 56.2%) and high-motion regions (5.9% vs. 22.7%).

Table 2. VLM prediction behavior on abnormal regions of selected and unselected events on UCF-Crime. Percentages are relative to abnormal regions within each motion group.

EventMotion# Abnormalprteach=1\boldsymbol{p^{\mathrm{teach}}_r=1} ↑\uparrowprteach=0\boldsymbol{p^{\mathrm{teach}}_r=0} ↓\downarrow
SelectedLow25132.7%25.9%
SelectedHigh62648.2%5.9%
UnselectedLow3156.7%56.2%
UnselectedHigh57622.6%22.7%

W2. Training Objective:

We thank the reviewer for highlighting the need to clarify the effective training objectives.

We have revised the Sec. 3.5 Training Objective section and Appendix B.1 to explicitly specify the event-dependent MIL and distillation weights, their normalization, and the objective for vanilla KD.

Specifically, the event-dependent parameter λe\lambda_e determines the MIL and distillation weights as (1−λe)(1-\lambda_e) and λe\lambda_e, respectively. Selected events use (λe=λs)=0.9(\lambda_e=\lambda_s)=0.9, while unselected events use (λe=λu)=0(\lambda_e=\lambda_u)=0, retaining MIL supervision without distillation. For indiscriminate (vanilla) KD, all abnormal events receive the same weight λe=0.9\lambda_e=0.9.

We thank the reviewer for maintaining a positive assessment of our work. We hope these additional analyses and clarifications adequately address your concerns. If you have any remaining questions or concerns, we would greatly appreciate the opportunity to address them.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.