cs.CVOct 1, 2026
SaveCog-VADU: A Training-Free Cognitive Reasoning Framework for Video Anomaly Detection and Understanding
Organizations: Centre for Vision, Speech and Signal Processing (CVSSP) University of Surrey, UK · Surrey Institute for People-Centred AI (PAI) University of Surrey, UK
Abstract
Video Anomaly Detection (VAD) aims to temporally localize abnormal events in videos. Most existing approaches rely on dataset-specific training and curated annotations, limiting generalization in open-set scenarios. Recent zero-shot methods based on Large Vision- Language Models (LVLMs) alleviate this dependency but often lack temporal continuity and structured reasoning. We propose Cog-VADU, a fully training-free framework that reformulates VAD as a sequential cognitive reasoning task. Cog-VADU introduces Chain-of- Anomaly Detection Thought Prompting (CoADTP), which unrolls an LVLM into a recurrent reasoning chain across video segments. By propagating structured rationales over time, the model maintains implicit temporal memory, enabling robust discrimination between com- plex anomalies and high-motion normal activities. To improve reliability, we further design a cross-modal re-ranking stage that aligns textual rationales with visual embeddings, enforcing semantic consistency and temporal coherence for refined and stable predictions. Extensive experiments on multiple public VAD benchmarks demonstrate that Cog-VADU achieves competitive zero-shot performance. Moreover, cross-model evaluations show that CoADTP consistently enhances reasoning-based anomaly detection in a model-agnostic manner, pro- viding interpretable and generalizable anomaly understanding for real-world applications.
Figures & tables
Figure 1: Comparison of video anomaly detection paradigms. (a) Vision-only methods rely on supervised training and produce anomaly scores without semantic reasoning. (b) Multimodal LLM-based VAD generates anomaly confidence and provides post-hoc explanations. (c) Cog-VADU employs a frozen LVLM with structured prompting and temporal cues to perform joint, training-free anomaly detection and reasoning.
Figure 2: Comparison of prior-aware prompts for zero-shot VAD.
Figure 3: Workflow of CoADTP integrating anomaly priors and temporal reasoning cues.
| Prompt | Prior | Reason | Temporal | AUC | AP | Acc@0.5 |
| Prompt 1 (Minimal) | ✗ | ✗ | ✗ | 0.7111 | 0.2150 | 78.28 |
| Prompt 2 (Format-Structured) | ✓ | ✗ | ✗ | 0.7403 | 0.2119 | 57.93 |
| CoADTP(Raw Stage) | ✓ | ✓ | ✓ | 0.7888 | 0.3773 | 82.41 |
Table 1: Ablation of prompting strategies in the RAW stage on UCF-Crime.Prompt 1 is a minimal caption-style prompt; Prompt 2 is a format-structured surveillance prompt with anomaly categories and an output template but no reasoning; CoADTP adds step-wise reasoning, an anomaly database, and temporal feedback.
Figure 4: Overview of Cog-VADU , a training-free, reasoning-driven video anomaly detection framework. Videos are divided into clips and processed by a frozen using Chain-of-Anomaly-Detection Prompting (CoADTP) to produce anomaly scores and textual rationales. A refinement stage employs ImageBind embeddings to retrieve and re-rank semantically related visual-text contexts, yielding final contextualized anomaly scores via softmax-weighted similarity aggregation.
Figure 5: UCF Sultani et al. (2019) Re-Annotation Examples. Case (a) shows missing anomaly regions in the original annotations. Case (b) highlights cause–and–effect anomaly segments previously ignored but included in our unified annotations.
| Method | BLEU | METEOR | ROUGE-L | ||||||
| C | E | V | C | E | V | C | E | V | |
| Training + Instruction-Tuned VAD MLLMs | |||||||||
| Hawk Tang et al. (2024) | 0.320 | 0.165 | 0.202 | 0.228 | 0.191 | 0.196 | 0.156 | 0.104 | 0.114 |
| Holmes-VAD Zhang et al. (2024b) | 0.514 | 0.318 | 0.306 | 0.224 | 0.237 | 0.211 | 0.235 | 0.164 | 0.161 |
| Holmes-VAU Zhang et al. (2025b) | 0.913 | 0.804 | 0.566 | 0.190 | 0.165 | 0.121 | 0.329 | 0.370 | 0.355 |
| Open-source MLLMs | |||||||||
Table 4: Comparison of hierarchical video anomaly reasoning performance across clip-level (C), event-level (E), and video-level (V) granularity. BLEU denotes cumulative BLEU-1 to BLEU-4 scores. Results demonstrate the effectiveness of reasoning-driven approaches for structured anomaly understanding from short clips to long videos. Bold means best and Underline means second best in each group . Results with + CoADTP show the impact of the Chain of Anomaly Detection Thought Prompt.
| Method | Judg. | Desc. | Anal. | Capt. | Overall |
| Training + instruction-tuned VAD MLLMs | |||||
| HAWK Tang et al. (2024) | 0.078 | 0.075 | 0.071 | 0.077 | 0.077 |
| Holmes-VAD Zhang et al. (2024b) | 0.167 | 0.236 | 0.235 | 0.263 | 0.245 |
| Holmes-VAU Zhang et al. (2025b) (SOTA, in-domain) | 0.652 | 0.297 | 0.479 | 0.430 | 0.431 |
| Open-source MLLMs + CoADTP (zero-shot) | |||||
| Video-ChatGPT Maaz et al. (2024) + CoADTP | 0.400 | 0.170 | 0.084 | 0.104 | 0.149 |
Table 5: BERTScore F1 Zhang et al. (2020) on HIVAU-70k Zhang et al. (2025b) (intersection of 350 IDs covered by all 10 methods; RoBERTa-large Zhang et al. (2020) , baseline-rescaled). Cog-VADU achieves training-free performance competitive with strong CoADTP-augmented backbones while substantially outperforming the training-based explainable baseline Holmes-VAD Zhang et al. (2024b) . Best in each column is in bold ; second-best is underlined .
| Method | SORA | OpenSORA | RG2 | VideoLCM | MS-T2 | Avenue | Ped1 | Ped2 |
| MLLM-based VAD (Training+ instruction Tuned ) | ||||||||
| Holmes-VAU Zhang et al. (2025b) | 2.17 | 34.00 | 24.00 | 29.81 | 25.00 | 6.06 | 3.33 | 5.56 |
| Holmes-VAD Zhang et al. (2024b) | 6.52 | 34.00 | 32.00 | 33.56 | 22.92 | 12.12 | 20.00 | 5.56 |
| HAWK Tang et al. (2024) | 24.64 | 52.00 | 44.00 | 36.54 | 50.00 | 36.36 | 36.67 | 38.89 |
| VAD-R1 Huang et al. (2025) | 41.30 | 78.00 | 56.00 | 63.46 | 60.42 | 75.76 | 60.00 | 63.89 |
| Open-Source MLLMs | ||||||||
Table 6: Performance comparison on VANE-Bench. Bharadwaj et al. (2025) Accuracy (%) across different video generation models and surveillance datasets. bold means best and underline means second best in a group.
Figure 6: Qualitative comparison on VANE. Cog-VADU generates structured, evidence-grounded reasoning (Observation Anomaly Reasoning Answer), correctly identifying the anomaly (B). In contrast, prior multimodal VAD methods, including RL-based VAD-R1 Bharadwaj et al. (2025) , often produce weakly grounded explanations.
Appendix figures & tables15 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8: UCF Sultani et al. (2019) and XD-Violence Wu et al. (2020) Re-Annotation Examples. Case (a) shows missing anomaly regions in the original annotations but included in our unified annotations ( orig.Anot: Original Annotations, Our. Anot: Ours Re-Annotations ). Case (b) highlights cause–and–effect anomaly segments previously ignored but included in our unified annotations.
Figure 9: Quantitative analysis of re-annotation changes under the proposed unified labeling protocol. (a) Original vs. corrected anomalous frame volumes. (b) Composition of re-annotated frames showing original and newly added portions. (c–d) Distribution of added frames across three identified annotation error types for XD-Violence Wu et al. (2020) and UCF-Crime Sultani et al. (2019) , respectively.
Figure 10: Structural analysis of annotation distributions under the proposed unified protocol. (a) Anomaly perception word cloud illustrating semantically enriched anomaly categories introduced through re-annotation, including subtle, contextual, and effect-based states that were under-represented in the original benchmarks. (b) Temporal relation word cloud highlighting causal, escalation, and aftermath relations, emphasizing the necessity of modeling complete cause-effect chains rather than isolated trigger events. Together, these visualizations provide linguistic and structural evidence that the original annotations are biased toward trigger-centric labeling, while our unified protocol captures temporally consistent and semantically grounded anomaly states.
Figure 11: Workflow of CoADTP integrating anomaly database and temporal context and reasoning cues.
| Configuration | Reasoning | Anomaly Database | Temporal Feedback | Refine- ment | AUC | AP |
| Base Prompt | ✗ | ✗ | ✗ | ✗ | 0.7111 | 0.2150 |
| + Reasoning only | ✓ | ✗ | ✗ | ✗ | 0.7548 | 0.2488 |
| + Anomaly DB only | ✗ | ✓ | ✗ | ✗ | 0.7406 | 0.3392 |
| + Temporal FB only | ✗ | ✗ | ✓ | ✗ | 0.7439 | 0.3303 |
| + Reasoning + DB | ✓ | ✓ | ✗ | ✗ | 0.7810 | 0.3238 |
| + R + DB + Temporal FB | ✓ | ✓ | ✓ | ✗ | 0.7888 | 0.3773 |
Table 9: Ablation study on UCF-Crime showing the contribution of each Cog-VADU component. We report (i) the base prompt, (ii) each textual component evaluated in isolation above the base prompt (rows 2–4, with score smoothing disabled), and (iii) cumulative configurations leading to the full pipeline (rows 5–7). Reasoning is the strongest AUC driver; Anomaly DB and Temporal FB are the strongest AP drivers; components are complementary rather than redundant.
Figure 13: Case 1 — Visually anomalous, semantically normal. Comparison of four state-of-the-art VAD methods on an open-web arm wrestling video (ground truth: Normal ). URDMU Zhou et al. (2023a) produces a strong false alarm (peak ); HolmesVAD Zhang et al. (2024b) hallucinates an unrelated “man in a blue shirt” scene; HAWK Tang et al. (2024) fabricates a basketball court context; and HolmesVAU Zhang et al. (2025b) contradicts its own reasoning by labelling recognized wrestling activity as “anomaly events.” In contrast, Cog-VADU correctly identifies the professional-wrestling context and outputs a calibrated score of ( Normal ).
Figure 14: Case 2 — Visually subtle, semantically anomalous. Comparison on an open-web video of the Tel Aviv missile interception (ground truth: Explosion ). URDMU Zhou et al. (2023a) barely registers the event (peak ); HolmesVAD Zhang et al. (2024b) hallucinates a “blue light in the sky”; HAWK Tang et al. (2024) confabulates a vehicle accident and “firework gone wrong”; and HolmesVAU Zhang et al. (2025b) misclassifies the projectiles as “fireworks or other sources of light.”
Figure 15: Failure case — categorical labelling error under graceful degradation. Cog-VADU is applied to an open-web video of a sensory room being flooded with water. The model produces a factually accurate description (“water starts flooding the room, causing objects to float and move around”), correctly identifies the reasoning (“not a normal activity and poses a safety concern”), and outputs a well-calibrated anomaly score of . However, the predicted anomaly type is Suspicious instead of the more precise Flood . The root cause is structural rather than reasoning-related: our predefined anomaly database does not contain a Flood category, and the model gracefully defaults to the generic Suspicious class. This failure is correctable in a training-free manner by simply extending the textual anomaly database, in contrast to opaque training-induced failures observed in baselines.
| Corruption Type | Recovery Rate | Median | Mean Max | Mean Final |
| Wrong category (Normal Violence) | 20/20 (100%) | 1 | 0.509 | 0.139 |
| Inverted (Anomaly Normal) | 20/20 (100%) | 1 | 0.446 | 0.180 |
| Hallucinated entity (Masked + Weapon) | 20/20 (100%) | 1 | 0.512 | 0.142 |
Table 12: Recovery statistics under three rationale corruption types on 20 UCF-Crime videos. All runs recover within the test horizon; the median recovery time is 1 segment for every corruption type.
Figure 16: Error propagation under injected rationale corruption. For each of three corruption types, we plot the mean absolute deviation between corrupted-run and baseline-run anomaly scores at each segment index after the injection point ( ). Shaded bands denote across 20 UCF-Crime videos. Deviations remain bounded ( – ) rather than accumulating upward, demonstrating that Cog-VADU’s per-segment visual grounding prevents long-range error propagation.
| Method | AP (Orig.) | AP (Ours-Ann.) |
| Weakly Supervised Non-Explainable Methods | ||
| UR-DMU Zhou et al. (2023a) [AAAI-23] | 35.59 | 50.84 |
| Explainable VAD Methods | ||
| VERA Ye et al. (2025) [CVPR-25] | 32.83 | 45.12 |
| Holmes-VAU Zhang et al. (2025b) [CVPR-25] | 36.24 | 56.49 |
| Training-Free Explainable Methods | ||
Table 13: Average Precision (AP %) on the UCF-Crime dataset Sultani et al. (2019) . “Orig.” denotes results evaluated on the original annotations, while “Ours-Ann.” reflects performance on our re-annotated split. Colored sections distinguish Weakly Supervised Non-Explainable , Explainable , and Training-Free Explainable methods.
Figure 17: Qualitative results of our Cog-VADU method on two test videos, from UCF-Crime Sultani et al. (2019) . For each video, plot the frame-wise anomaly scores and show keyframes with their most relevant temporal summaries (green for normal, red for abnormal) . These examples illustrate alignment between predicted anomaly scores, visual content, and descriptions. Ground-truth anomalies are highlighted in red.
Figure 18: Qualitative results of our Cog-VADU method on two test videos, from XD-Violence Wu et al. (2020) . For each video, plot the frame-wise anomaly scores and show keyframes with their most relevant temporal summaries (green for normal, red for abnormal) . These examples illustrate alignment between predicted anomaly scores, visual content, and descriptions. Ground-truth anomalies are highlighted in red.
Figure 19: Qualitative comparison on a video reasoning benchmark. Our training-free framework Cog-VADU (blue) correctly identifies the anomaly by reasoning over temporal visual inconsistencies, detecting that the second puppy’s nose and tongue morph unnaturally across frames. In contrast, training-based VAD approaches (orange), including HAWK Tang et al. (2024) , HolmesVAU Zhang et al. (2025b) , and VAD-R1 Huang et al. (2025) , either hallucinate unrelated changes (e.g., ear color or fur texture) or fail to localize the correct anomaly despite similar visual evidence. This example highlights the advantage of reasoning-driven anomaly detection in capturing subtle semantic inconsistencies that are not easily modeled by conventional training-based approaches.