Paper deep dive
O-VAD: Industrial Video Anomaly Detection through Object-Centric Tracking and Reasoning
Mei Yuan, Qi Long, Qifeng Wu, Zhenyang Li, Yizhou Zhao, Lei Wang, Yang Liu, Min Xu
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 7/21/2026, 6:21:12 AM
Summary
The paper introduces O-VAD, a training-free agentic framework for Industrial Video Anomaly Detection (IVAD) that utilizes object-centric tracking and reasoning. Unlike traditional methods requiring retraining or domain-specific knowledge, O-VAD employs a 'ground-track-reason' pipeline using Vision-Language Models (VLMs) to track spatial-temporal dynamics of objects, detect state changes, and generate interpretable anomaly reports with causal analysis.
Entities (12)
Relation Signals (10)
Mei Yuan → affiliatedwith → Carnegie Mellon University
confidence 95% · Mei Yuan 1 ... 1 Carnegie Mellon University
O-VAD → solves → Industrial Video Anomaly Detection
confidence 95% · O-VAD: Industrial Video Anomaly Detection through Object-Centric Tracking and Reasoning
O-VAD → uses → Vision-Language Models
confidence 92% · we introduce a training-free agentic framework... emphasizing object state evolution... VLM-based anomaly reasoning
O-VAD → evaluatedon → IPAD
confidence 90% · Extensive experiments on three IVAD datasets... IPAD
O-VAD → evaluatedon → Phys-AD
confidence 90% · Extensive experiments on three IVAD datasets... Phys-AD
O-VAD → evaluatedon → LiquidAD
confidence 90% · Extensive experiments on three IVAD datasets... LiquidAD
O-VAD → usescomponent → SAM3
confidence 90% · Objects are first detected and segmented via VLM-grounded masking with SAM3
O-VAD → usescomponent → SAM2
confidence 90% · tracked forward via SAM2 to form tubelets
O-VAD → →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Industrial Video Anomaly Detection (IVAD) aims to identify anomalous objects and events in an industrial process, which is crucial for modern manufacturing and quality control systems. Existing VLM-based anomaly reasoning methods are capable of detecting open-ended anomalies in general domains. However, their performance declines in industrial settings characterized by intricate object transformations, strict physics, and procedural constraints. To tackle the complexity of such interaction-intensive detection, we introduce a training-free agentic framework for anomaly detection free of domain-specific knowledge, emphasizing object state evolution like humans inspectors. It is designed to track spatial-temporal dynamics and underlying transformations of detected objects over time, and then reason over the object-wise temporal state trajectories to identify abnormal objects in grounded frames. Our method overcomes limitations of prior approaches that rely on retraining on normal clips or injecting domain knowledge as context for test-time inference. Extensive experiments on three IVAD datasets demonstrate that our method outperforms frontier VLMs, agentic frameworks, and traditional VAD methods fine-tuned on the respective datasets, while providing interpretable reports over anomaly processes and types.
Tags
Links
- Source: https://arxiv.org/abs/2607.18142v1
- Canonical: https://arxiv.org/abs/2607.18142v1
PDF not stored locally. Use the link above to view on the source site.
Full Text
177,450 characters extracted from source content.
Expand or collapse full text
O-VAD: Industrial Video Anomaly Detection through Object-Centric Tracking and Reasoning Mei Yuan 1 , Qi Long 1 , Qifeng Wu 1 , Zhenyang Li 2 , Yizhou Zhao 1 , Lei Wang 3 , Yang Liu 1 , and Min Xu 1⋆ 1 Carnegie Mellon University 2 University of Alabama at Birmingham 3 Griffith University meiyuan@andrew.cmu.edu, mxu1@cs.cmu.edu https://o-vad.github.io/ O-VAD (Ours) Answer: Yes, Toothpaste Tube (Darlie), leakage,170-179frames, causal analysis. "report": “overall_anomaly”: ”yes“, ”anomaly_type“:“material_anomaly/unintended_extrusion_or_leakage", "objects": [ "name": "Toothpaste Tube (Darlie)", "anomaly_segments": [[170,179]], "anomaly_object_ids": ["0"], ”state tracking": "A surface puncture or breach appears on the tube starting around frame 170, indicating unintended failure despite controlled compression.", “evidence”: [“”start_frame“: 170, ”end_frame“: 179, ”change_type“: ”surface_change“, ”description“: ”In the later frame, a small puncture or opening has appeared on the upper surface of the tube (near the center, under where the pointed tool was positioned in the earlier frame). This is visible as a small, bright spot that was not present before, indicating a breach in the tube‘s surface.“, ”severity“: ”slight“, ”object_idx“: ”0”"], Query : "Youareanexpertindustrialanomalydetectionagent. Yourgoalistoanalyzethisvideotodetectanomalies inthevideo.Pleasestartwithastep-by-stepanalysis oftheframes,thengeneratetheanomalyreport." w/o training FeatureEmbedding-basedVAD Answer:1 vs0 w/ training predefined labels GPT-5 Answer: “No anomaly detected.” "report": "is_anomalous": "No", "anomaly_type": "None", "reason": ”<think>The toothpaste tube undergoes normal deformation under controlled pressure from the mechanical clamp without any signs of leakage, breakage, or malfunction. </think><answer>No anomaly detected</answer>” w/o training semanticrationale open-ended labels Reconstruction-based VAD Answer:1 vs0 w/ training predefined labels state trajectoriessemanticrationaleopen-ended labels raw video |------------------------------------------state1: pressing----------------------------------------|--------------------------------------state2: deformation----------------------------------|-------------------------------state2: leakage----------------------------------|---------- Fig. 1: Why object state evolution matters for industrial anomaly detec- tion. The toothpaste tube undergoes pressing and deformation before a subtle leakage emerges. Traditional methods predict binary classifications. VLM prompting fails to detect it due to the lack of evidence. O-VAD tracks object-wise state changes and pro- duces an open-ended anomaly report with in-depth analysis. Abstract. Industrial Video Anomaly Detection (IVAD) aims to identify anomalous objects and events in an industrial process, which is crucial for modern manufacturing and quality control systems. Existing VLM-based anomaly reasoning methods are capable of detecting open-ended anoma- lies in general domains. However, their performance declines in industrial settings characterized by intricate object transformations, strict physics, and procedural constraints. To tackle the complexity of such interaction- intensive detection, we introduce a training-free agentic framework O- VAD for anomaly detection free of domain-specific knowledge, empha- sizing object state evolution like humans inspectors. It is designed to ⋆ Corresponding author. arXiv:2607.18142v1 [cs.CV] 20 Jul 2026 2M. Yuan et al. track spatial-temporal dynamics and underlying transformations of de- tected objects over time, and then reason over the object-wise temporal state trajectories to identify abnormal objects in grounded frames. Our method overcomes limitations of prior approaches that rely on retraining on normal clips or injecting domain knowledge as context for test-time inference. Extensive experiments on three IVAD datasets demonstrate that our method outperforms frontier VLMs, agentic frameworks, and traditional VAD methods fine-tuned on the respective datasets, while providing interpretable reports over anomaly processes and types. Keywords: Industrial Anomaly Detection· Agentic Reasoning· Vision- Language Models 1 Introduction Industrial Video Anomaly Detection (IVAD) aims to identify anomalous ob- jects and events in an industrial video process. It plays a critical role in modern manufacturing and quality control systems, where automated visual inspection can significantly reduce costs and improve production reliability [26]. Industrial processes are extremely challenging and different from general domains, since they are usually characterized by: (i) complex object transformations: objects undergo substantial physical and functional state changes through cutting, as- sembly, heating, pressing, and other manufacturing operations; (i) strict physical and procedural constraints: anomalies manifest as violations of expected spatial configurations, periodic procedures, or physical laws; (i) high interpretability re- quirements: operators need explicit explanations of detected anomalies for root cause analysis and corrective actions. Traditional Video Anomaly Detection (VAD) methods primarily adopt two paradigms: reconstruction-based approaches [8,9,12,42–44] reconstruct anoma- lous samples to their corresponding normal counterparts and calculate the recon- struction error, and embedding-based [6,11,15,17,32] methods focus on model- ing the feature embeddings of normal samples and measure deviations. But they typically follow the “one-class-one-model” learning paradigm, requiring plentiful normal samples for each object class to learn its distribution [10], which makes it impractical for industrial anomaly detection settings and less suitable for dy- namic production environments. Given that traditional VADs have limited generalization abilities and binary classifications provide no semantic rationale for their decisions, recent works [10,14,18,24,25,45,48,49] propose a series of VLM-based anomaly detection methods, featuring semantic anomaly understanding and test-time detection inference. Although VLMs have unlocked their potential for detecting anomalies aug- mented with semantic-visual understanding in general domains, there are still two issues that remain up in the air: (1) Their performance in industrial settings declined due to lack of domain-specific knowledge and fine-grained annotations. Recent training-free studies [16] provide external knowledge or normal samples in context for test-time inference to alleviate this limitation. However, they O-VAD3 heavily rely on sophisticated prompt design to caption videos, inject knowledge, and then assign anomaly scores by off-the-shelf VLMs. The overly dependence on external context leads to misaligned responses that prior- itize contextual plausibility over accuracy [18]. (2) Video/frame level features or descriptions have weak understanding of localized details within objects over time, which hinders their reliable and accurate interaction-intensive IVAD reasoning. Some VLM-based methods exploited the spatio-temporal fusion [20,27,34,38,39], and detect anomalies from when and where perspectives [13,24]. However, they lose focus on object-centric physics and interactions over time. It is the core when human inspectors detect object-centric state changes throughout industrial processes, especially when objects undergo substantial physical and functional state changes. To tackle these two issues, we propose O-VAD, a training-free agentic frame- work that detects anomalies by tracking object state evolution over time—free of any domain-specific knowledge, labels, or predefined anomaly taxonomy. (1) For the first issue, rather than injecting external knowledge or relying on com- plex prompt design, O-VAD enables the VLM to build its own understanding through a “ground→track→reason” agentic pipeline. Objects are first detected and segmented via VLM-grounded masking with SAM3 [2], then continuously tracked and queried across frames to accumulate structured evidence. A cascaded chain-of-thought (CoT) then derives anomaly judgments from the VLM’s own internalized commonsense and physical knowledge, mirroring the cognitive pro- cess of human inspectors. (2) For the second issue, we shift the unit of analysis from holistic frames to individual objects. An object state tracker maintains per-object representations and queries the VLM at key temporal transitions to capture fine-grained state changes, e.g. deformation, material release, surface al- teration. These structured state trajectories ground the downstream reasoning, making detection both object-centric and spatio-temporally aware, and yielding open-ended, fine-grained outputs: anomaly types, grounded abnormal frames, affected objects, and causal analyses. Result-wise, despite using no labor-intensive annotations, the proposed agen- tic framework delivers state-of-the-art performance on any level anomaly de- tection, surpassing frontier VLMs and agentic frameworks on quantitative and qualitative evaluations. To summarize our contributions: (1) We present a generalizable data curation pipeline and apply it to three object-centric IVAD datasets (IPAD, Phys-AD and LiquidAD) with fine-grained annotations of object trajectories across frames. (2) We propose a training-free, three-stage agentic reasoning framework that requires no domain-specific knowledge or fine-tuning, performing object discov- ery, spatiotemporal tubelet construction withy open-ended state change detec- tion, and chain-of-thought anomaly reasoning with visual verification. (3) Our framework achieves state-of-the-art performance on three IVAD datasets at both video- and frame- levels, outperforming frontier VLMs (Qwen3- VL-32B and GPT-5), two agentic framework for VAD, and fine-tuned traditional 4M. Yuan et al. AD methods, while providing interpretable reasoning over anomaly processes and open-ended types. 2 Related Work Traditional Video Anomaly Detection. Traditional video anomaly detec- tion methods primarily adopt two paradigms: reconstruction-based approaches [8, 9,12,42–44] flag anomalies by their elevated reconstruction error, while embedding- based methods [6,11,15,17,32] model the feature distribution of normal samples and measure deviations from it. Despite strong benchmark performance, both share three limitations. First, they follow a “one-class-one-model” paradigm [10], requiring abundant normal samples and a separate model per object class, and must be re-trained for every unseen domain or anomaly type [22], mak- ing them impractical for novel categories and dynamic production lines. Sec- ond, they are largely black-box, emitting an anomaly score without semantic rationale [39], which has motivated growing interest in explainable, language- grounded anomaly understanding [22, 46]. Third, they compress entire frames into holistic features, neglecting region-level cues and over-relying on dominant background context [21]. Our O-VAD addresses these limitations with a training- free, object-centric framework that requires no per-class training, produces se- mantic explanations for its decisions, and reasons over region-level evidence. Multimodal Video Anomaly Detection Recent works [10, 14, 18, 24, 25, 45, 48,49] revisit anomaly detection with VLM-based methods, valued for their gen- eralizability and explainability, especially in the zero/few-shot setting. These approaches rely on sophisticated prompt design that captions videos and assigns anomaly scores via off-the-shelf VLMs. To strengthen spatio-temporal aware- ness, some methods couple this with auxiliary temporal modeling or prompt- ing [20,27,34,38,39]; for instance, a unified zero-shot framework [24] chains tem- poral detection, spatial localization, and textual explanation through a test-time reasoning process over foundation models, requiring no additional training. Yet these methods largely overlook that superficial temporal caption sequences lose focus on the object-centric dynamics and interactions over time—precisely the cues human inspectors rely on to detect anomalous objects and events through- out industrial processes. In contrast, our O-VAD explicitly models object-centric spatio-temporal evidence while retaining the training-free and explainable ad- vantages of VLM-based detection. 3 Method Overview. Given an industrial process video V = I t T t=1 , O-VAD produces a structured anomaly report for each detected anomaly, it outputs the anomaly type, severity, affected object, temporal localization, and a natural-language causal explanation—all without any domain-specific knowledge, predefined tax- onomy, or training data. To achieve this, O-VAD adopts a three-stage agentic O-VAD5 (b) Object-Centric State Tracking spatiotemporal partitions state changes detection sate trajectory understanding (a) Object Grounding Q : Analyze this video and list ALL distinct objects/items visible. A : object names, natural-language description (material, initial state) sampled frames SAM mask prompt object metadata VLM Q : Compare these two frames from a video. The same object is highlighted with init_c_name contour in both. Carefully check for ANY of changes. Q : Analyze the interaction between objects (Object 1 with init_c_name contour and Object 2 withquery_c_namecontour) in this image. Describe the interaction. A : The blue object appears slightly repositioned upward and better aligned. A : The blue box appears to be standing on top of the green platform, suggesting a stabilizing or settling motion. VLM Time = 0s Time = 2s Time = 3sTime = 6s CF +SAM input frames &prompt objects VLM A : -Frame-by-frame masks, -affected object metadata -state change events spatiotemporal dynamics ❓ what ❓ when ❓ how ❓ where ❓ why ❓ proximity✅ semantics✅ (c) State-Aware AnomalyReasoning VLM process understanding. observation **Anomaly Detection report** -Video-level anomaly + anomaly types -Frame-level anomaly + anomaly frames -Object-level anomaly + anomaly states expectation comparison causation Classification & Severity post-verify anomaly frames Fig. 2: Overview of O-VAD. (a) Stage 1 samples multiple frames and queries a VLM to discover all objects, then segments them with SAM to produce initial masks and metadata. (b) Stage 2 constructs spatiotemporal tubelets via CropFormer [29] and SAM2 [31], recovers missing tracks through semantic and proximity priors, and query the VLM for open-ended state change detection and inter-object interaction analysis. (c) Stage 3 feeds the accumulated object metadata and state change events into a multi-step chain-of-thought reasoning chain that distinguishes process actions from failure outcomes, followed by visual verification of candidate anomalies. pipeline (Fig. 2). Stage 1 (§3.2) discovers and segments all task-relevant ob- jects via VLM-grounded masking with SAM [2], rather than manual annotation. Stage 2 (§3.3) tracks each object through potential state trajectories via spa- tiotemporal partitioning, querying the VLM at temporal transitions to detect fine-grained state changes and produce per-object state trajectories. Stage 3 (§3.4) reasons over the accumulated evidence through a multi-step chain-of- thought that separates abnormal stage changes from expected process outcomes, yielding anomaly judgments grounded in the VLM’s own commonsense and phys- ical reasoning. 3.1 Problem Formulation Our O-VAD decomposes the problem into an object-centric formulation. We first detect K objectsO =o k K k=1 with masksm (t) k via VLM-grounded segmenta- tion, then track each object’s state trajectoryS k =(m (t) k ,φ (t) k ) T t=1 , where m (t) k is the mask of object k at frame t and φ (t) k is a VLM-generated natural-language description of its physical state (e.g., name, material, surface condition, and detected state changes). Object-level anomaly reasoning is then performed as: ˆy t,k = θ VLM p CoT ⊕S k ⊕ c ,ˆy frm t = max k ˆy t,k ,ˆy vid = max t ˆy frm t ,(1) 6M. Yuan et al. where ˆy t,k ∈ [0, 1] is the anomaly confidence for object k at frame t, c is the optional video caption, p CoT is a cascaded chain-of-thought prompt that guides the VLM through cognitive anomaly reasoning, and ⊕ denotes concatenation of the prompt context. The frame- and video-level scores are obtained by max- pooling over objects and frames, respectively. By conditioning on per-object state trajectories S k rather than raw frames alone, O-VAD produces fine-grained, open-ended anomaly predictions with grounded frames, affected objects, and causal explanations; the discrete anomaly set A used for reporting is obtained from these scores after the reasoning and verification of Stage 3 (§3.4). 3.2 Stage 1: Automated Object Grounding Rather than relying on a single reference frame, we sample a set of candidate frames I t j N f j=1 spanning the video so that objects initially occluded or absent can still be discovered. For each candidate frame I t j , a VLM generates a struc- tured object inventory: O t j = φ VLM (I t j ) = (n k , d k , b k ) N t j k=1 , (2) where n k is the object name, d k a natural-language description (material, initial state), and b k a spatial cue. Inventories across frames are merged into a dedu- plicated set O = S j O t j . Each object o k ∈O is then segmented in a reference frame I 1 using SAM3 [2]: m (1) k = SAM(I 1 , o k ), M 1 = m (1) k |O| k=1 .(3) The output of Stage 1, including initial masks M 1 together with object meta- data, grounds all subsequent tracking and reasoning in concrete, per-object evi- dence. 3.3 Stage 2: Object-Centric State Tracking The second stage tracks each object across all frames, detecting and describing state changes even when objects undergo significant transformations. We build upon TubeletGraph [35] for tracking through transformations, and extend it with two novel components: object interaction analysis and open-ended state change annotation. Spatiotemporal Partitioning and Tubelet Construction. We construct a spatiotemporal partitionP of the video that associates every pixel region across all frames with a tracked entity. An entity segmentation model (CropFormer [29]) produces per-frame spatial partitions E t = CF(I t ), and each entity e i 1 ∈ E 1 (including the initial object masks M 1 ) is tracked forward via SAM2 [31] to form tubelets: P init =P i |E 1 | i=1 , P i =e i t T t=1 .(4) New tubelets are incrementally added whenever track-less regions emerge. For each entity ˆe j t ∈E t at frame t > 1, a new tubelet is initiated if less than τ coverage of its area is covered by existing tubelets. O-VAD7 State Tracking. To recover missing object tracks after state transformations (e.g., an object breaking apart or releasing material), candidate tracks that emerge after the initial frame are evaluated using spatial proximity and semantic consistency priors. Spatial Proximity. For a candidate track C =c s ,c s+1 ,...,c T beginning at frame s and the prompt object track P =p 1 ,p 2 ,...,p T , the spatial proximity score is: S prox (C,P) = max j∈1,2,3 |c s ∩ m j s | |c s | ,(5) where m j s 3 j=1 are the three candidate masks from SAM2’s multi-mask output at frame s. A candidate is considered proximal if S prox (C,P) > τ prox . Semantic Consistency. For a mask M and frame I, we compute the masked CLIP [30] feature f(M,I) = Pool(CLIP(I),M) via mask-pooling. The semantic similarity is: S sem (C,P) = max i∈1,...,s−1, j∈s,...,T f(p i ,I i )· f(c j ,I j ) ⊤ . (6) The final set of valid continuation tracks is: V =C ∈P | C begins at t > 0, S prox (C,P) > τ prox , S sem (C,P) > τ sem .(7) The complete tracking result T = P ∪V captures the object through transfor- mations. State Change Detection and Understanding. For object-aware frame pairs, the VLM is prompted with masked visualizations of both frames and asked to identify state changes through visual understanding. Critically, we adopt an open-ended labeling scheme: both the change type (e.g., compression deforma- tion, material release, rotational resistance) and the change cause (e.g., object leakage, environmental lighting) are generated as free-form natural language by the VLM. This enables the system to describe novel phenomena for generaliz- ing across diverse industrial processes. Each detected state change event e is represented as a tuple: e = (t start , t end , type, cause, desc, sev, k),(8) where t start and t end are the bounding frame indices, type is the free-form change type, cause is the causal label, desc is a natural-language description, sev ∈ none, slight, moderate, severe is the severity, and k is the affected object index. 3.4 Stage 3: State-Aware Anomaly Reasoning We formulate cognitive anomaly reasoning as a cascaded chain-of-thought (CoT) task over the accumulated object states. The final stage synthesizes the tracking metadata, state change events, and visual evidence into an anomaly detection 8M. Yuan et al. decision via multi-step reasoning. The key design principle is open-ended clas- sification: the VLM is not constrained to a fixed taxonomy but reasons freely from its domain knowledge, grounded in the tracking evidence. Finally, we have both the anomaly set and the complete reasoning trace for full interpretability. Cognitive Anomaly Reasoning. The core anomaly reasoning proceeds through a cascaded chain-of-thought (CoT) that mirrors how a human quality inspector diagnoses failures. The VLM receives, in a single call, the object metadata, fil- tered state changes, inferred task context, video caption, and sampled frames, and is instructed to reason through sequential steps with confidence scores. 1. Process Understanding. The VLM identifies the industrial process and its purpose, then triages each state change into expected process, potential anomaly, and normal noise. 2. Observation. The VLM cites specific object IDs, frame ranges, change types, and severities from the tracking data, focusing on events triaged as potential anomalies. 3. Expectation. The VLM articulates the pass/fail criteria for the identified test, distinguishing expected mechanical responses from failure-indicating outcomes. 4. Comparison. For each candidate anomaly, the VLM determines whether it represents a genuine failure outcome or a mechanical response that was mis- categorized. 5. Causation. The VLM reasons about root causes without being constrained to predefined categories. 6. Classification & Severity. A free-form anomaly type is assigned and each confirmed anomaly receives a severity rating based on its impact on task completion, safety implications, and reversibility. The reasoning chain outputs a set of candidate anomalies a, where each a carries a free-form anomaly type, an affected object index, a frame range, a severity level, and an initial confidence c orig (a) ∈ [0, 1] produced by the VLM. These candidates are then refined by the verification step below to form the final anomaly set A. Post Visual Verification. To suppress false positives, each candidate anomaly a is gated by its initial confidence c orig (a): anomalies with c orig (a) > τ hi bypass verification (they are sufficiently certain), those with c orig (a) < τ lo are discarded outright, and only the remaining candidates are verified against their evidence frames: (verified, c ver ) = φ VLM I t t∈evidence , c, a .(9) The verification step explicitly checks whether the claimed anomaly is actually a normal process behavior visible in the caption c or frames. The final confidence is then computed via a multiplicative gating scheme: c final (a) = ( c orig (a)· c ver if verified = true, c orig (a)· (1− c ver ) if verified = false, (10) O-VAD9 and the anomaly is retained only if c final (a)≥ τ conf . We use τ hi = 0.8, τ lo = 0.2, and τ conf = 0.3 in all experiments. The final anomaly set A and the complete reasoning trace together form the output reportR, providing full interpretability. 4 Experiment 4.1 Experiment Setup. Datasets & evaluation metrics. We evaluate on the official test splits of three industrial benchmarks spanning diverse anomaly types and interaction modalities. (i) Phys-AD [23]: a large-scale, physics-grounded dataset from a real robot arm and motor, with 22 object categories and 47 defect types across 6,434 videos (60 FPS); its anomalies require physical reasoning over interactions such as pressing, rotating, and grasping. We report per-category video-level AUROC, aggregate Acc/P/R/F1 at the video and anomaly-type levels, and BERTScore for anomaly-type semantic alignment. (i) LiquidAD [5]: liquid-transfer anoma- lies in automated labs, 2,251 videos (30 FPS) of 8 parallel pipettes dispens- ing into 8 test tubes; we report video- and frame-level Acc/P/R/F1/AUROC. (i) IPAD [26]: industrial-manufacturing processes (∼2,000 clips over 16 de- vice types, synthetic and real) that define anomalies as deviations from refer- ence normal samples rather than physical commonsense; we evaluate under this reference-based protocol and report video- and frame-level metrics. For VLM- based methods, an LLM-as-judge protocol assesses whether a predicted anomaly type semantically matches the ground truth, enabling fair comparison across heterogeneous outputs. Full dataset details and per-scenario results are in the supplement (§A.1 and §B.2). Hyperparameters & experiment details. O-VAD is fully training-free; un- less noted, we use the following defaults. Stage 1 (VLM-grounded masking) uses GPT-5 [33] as the VLM and SAM3 [2] for concept-aware zero-shot segmentation, with a permissive detection threshold of 0.1 to maximize discovery recall. Stage 2 (object state tracking) partitions each video into tubelets via CropFormer [29] for per-frame entity segmentation and SAM2 [31] for temporal propagation, sampling every 10th frame; track recovery is gated by spatial-proximity and semantic-consistency thresholds. Stage 3 (cognitive anomaly reasoning) applies the 6-step CoT prompt over the accumulated state trajectories, key frames, and object metadata to produce the structured report, after which confidence-tiered visual verification suppresses false positives. Phys-AD videos are subsampled with dynamic FPS to reduce redundancy while preserving key state transitions. All experiments run on a single NVIDIA H200 GPU (141 GB VRAM). Full threshold hyperparameters and model-selection rationale are in the supplement (§A.2 and §A.6). Baseline Methods. We compare O-VAD against three groups. (i) Traditional VADs: the best method from each paradigm on Phys-AD—MNAD.p [28] (unsu- pervised, prediction-based memory normality) and S3R [37] (weakly-supervised 10M. Yuan et al. Table 1: Video-level, type-level, and frame-level Acc/P/R/F1/AUROC (%) on Phys-AD [23], LiquidAD [5], and IPAD [26]. Methods are grouped into traditional VADs ([Trad.]), direct-prompting VLMs ([Direct Prompting]), agentic workflows ([Workflow]), and our [O-VAD]. Type-level BERTScore and LLM-judge scores are reported only for methods that emit free-form anomaly-type descriptions. [Trad.][Direct Prompting] [Workflow] [O-VAD(Ours)] Dataset Level MetricMNAD.p † S3R † Qwen3 ∗ GPT-5 ∗ URF ∗ VERA † O-VAD ∗ Phys-AD video Acc0.635 0.591 0.4540.6400.329 0.4700.592 P0.7130.739 0.7170.6900.853 0.5000.724 R 0.8030.645 0.3680.5800.053 0.0500.625 F10.7550.689 0.4860.6300.100 0.0910.621 AUC0.4950.555 0.5130.5020.426 0.4560.584 type BERTscore–0.7980.878–0.803 LLM-judge–0.3720.580–0.595 LiquidAD video Acc0.184 0.631 0.1500.4620.903 0.0910.868 P1.0000.966 0.8860.8960.912 0.0000.910 R0.102 0.616 0.0750.4620.988 0.0000.948 F10.185 0.752 0.1390.6100.949 0.0000.929 AUC0.4530.651 0.4890.4650.365 0.5340.692 frame Acc0.489 0.6010.8760.6330.5640.7090.458 P0.3950.800 0.0000.2530.602 0.0000.431 R0.589 0.590 0.0000.1340.708 0.0000.614 F10.472 0.679 0.0000.1760.651 0.0000.507 AUC0.4870.625 0.4990.4840.506 0.5000.512 IPAD video Acc0.555 0.545 0.5310.6070.5820.6580.582 P0.6430.836 0.5810.3750.590 0.2380.588 R0.529 0.271 0.7000.3860.929 0.0890.954 F10.581 0.409 0.6350.3880.721 0.1300.714 AUC0.5220.539 0.4980.5190.417 0.5190.565 frame Acc0.716 0.661 0.7030.8150.6710.8520.515 P 0.410 0.309 0.1140.2730.331 0.0000.291 R0.0060.166 0.0030.1190.164 0.0000.477 F10.012 0.216 0.0060.1650.220 0.0000.338 AUC0.430 0.495 0.4970.5320.512 0.4900.518 “†” = trained; “∗” = training-free. Colored cells = best; lighter ones = 2nd best. sparse representation); both require training data and produce only binary scores without explanations. (i) Direct Prompting: frontier VLMs in a zero-shot “caption- then-classify” setup—Qwen3-VL-32B [1] (extended thinking) and GPT-5 [33]— receiving raw frames and a standard anomaly-detection prompt, without object grounding or state tracking. (i) Agentic Workflows: two recent training-free frameworks. URF-ZS-HVAA [24] chains temporal detection, spatial localization, and textual explanation via intra- and inter-task reasoning; VERA [40] verbal- izes learnable guiding questions for a frozen VLM, decomposing the task into yes/no perceptual queries aggregated into a video-level verdict. 4.2 Quantitative Results Overall comparison. Table 1 summarizes aggregate Acc/P/R/F1/AUROC across the three benchmarks. Despite using no training data, domain knowledge, or predefined taxonomies, O-VAD attains the best average video-level AUROC O-VAD11 Table 2: Video-level AUROC (↑) on Phys-AD [23] across 22 object cat- egories. “†” = methods requiring training data. Per category, the best / 2nd-best across all methods is /; our best results are additionally in bold. Trad.Direct Prompting Workflow O-VAD (Ours) CategoryMNAD.p † S3R † Qwen3 ∗ GPT-5 ∗ URF ∗ VERA † O-VAD ∗ Ball0.589 0.155 0.5280.744 0.461 0.5480.579 Button0.526 0.108 0.5060.4560.717 0.4690.393 Car 0.599 0.527 0.4920.617 0.497 0.3830.575 Caster Wheel 0.4680.830 0.5330.543 0.4100.6090.301 Clip0.6280.537 0.4420.518 0.275 0.1800.423 Clock0.513 0.529 0.581 0.766 0.483 0.5000.669 Fan0.4090.602 0.5660.472 0.523 0.3890.669 Gear0.418 0.358 0.6440.397 0.451 0.5120.879 Hinge0.4970.7160.5890.516 0.544 0.0410.526 Liquid0.453 0.7330.8330.512 0.6910.8270.776 Lock0.442 0.323 0.3880.231 0.0290.5570.780 Magnet0.512 0.724 0.5920.364 0.5180.8920.576 Roll. Bear.0.437 0.016 0.4180.451 0.3040.5000.650 Rub. Band0.561 0.4630.7670.2790.783 0.2760.721 Screw0.638 0.620 0.4500.441 0.298 0.4000.704 Servo0.479 0.4520.4920.330 0.342 0.3410.495 Slide0.5340.533 0.5290.528 0.345 0.2250.415 Sph. Bear.0.547 0.894 0.4830.440 0.629 0.1140.641 Sticky Roller 0.5170.778 0.5330.523 0.713 0.4740.811 Toothpaste0.257 0.535 0.5780.832 0.528 0.5790.689 U Disk0.5130.585 0.5780.437 0.404 0.3110.606 Zipper0.2620.7910.5170.477 0.273 0.2240.388 Avg.0.4950.555 0.5130.503 0.426 0.4560.584 among training-free methods on Phys-AD (0.584), beating the direct-prompting baselines Qwen3-VL-32B (0.513) and GPT-5 (0.503) and the agentic workflows URF-ZS-HVAA (0.426) and VERA (0.456), and even exceeding both trained baselines MNAD.p (0.495) and S3R (0.555). It also yields the most semanti- cally faithful descriptions, with the best type-level BERTScore (0.803) among training-free methods. The same pattern holds on LiquidAD and IPAD. This shows the bottleneck in VLM-based IVAD lies not in language reasoning but in object-level evidence construction: O-VAD’s “ground→ track→ reason” pipeline avoids the false-negative collapse seen in caption-level baselines. Per-dataset analysis. Table 2 reports per-category video-level AUROC on Phys-AD across all 22 object categories. O-VAD claims the best or second-best AUROC on 16 of 22 categories, with the largest gains on categories involving complex multi-step interactions (e.g., Gear 0.879, Lock 0.780, Screw 0.704) where temporal state tracking is most critical, and on those with subtle surface-level anomalies (e.g., Sticky Roller 0.811, U Disk 0.606) where object-centric analysis 12M. Yuan et al. reveals details invisible to frame-level approaches. The other two benchmarks stress complementary capabilities. LiquidAD requires per-instance precision in a cluttered scene with up to eight simultaneous pipettes: zero-shot VLMs collapse to near-zero recall and a strong bias toward “normal,” whereas O-VAD tracks per-pipette state trajectories to attain the best training-free video-level AUROC (0.692) and the strongest frame-level recall and F1. IPAD shifts from physics- grounded reasoning to reference-based deviation across 16 industrial scenarios, where O-VAD again achieves the highest average video-level (0.565) and frame- level (0.518) AUROC and ranks first on 12 of 16 scenarios at the video level. Together these results show that object-centric state tracking generalizes across multi-instance laboratory scenes and periodic industrial processes alike. Per- category breakdowns reporting all five metrics (Acc/P/R/F1/AUROC) for every category and scenario on all three datasets are provided in the supplementary material (§B.1 and §B.2). 4.3 Qualitative Results To illustrate how object-centric evidence shapes reasoning quality, we compare O-VAD against GPT-5 [33]—prompted with the same cascaded chain-of-thought but without object grounding or state tracking—on three representative Phys- AD scenarios (Figure 3). Additional reasoning traces, failure-case analyses, and human and LLM-as-judge studies are provided in the appendix (§C). Case Analysis. Figure 3 visualizes full reasoning traces on representative Phys- AD examples. In Case 1 (plastic bottle rotation), a clamped bottle leaks severely at frames 60–80. GPT-5 reports “no leakage,” while O-VAD tracks progressive deformation (frames 0–60), detects material release at frame 60, and classifies a high-severity loss-of-containment anomaly. In Case 2 (hinge screw fastening), three defects co-occur: repeated screw back-out (frames 30–110), head-slot strip- ping (frames 80–90), and screwdriver shaft bending (frames 60–100). GPT-5 con- cludes normal tightening; O-VAD captures all three streams and attributes them to torque-control or bit-mismatch failure. Across all cases, GPT-5 defaults to “no anomaly” without object state evidence, whereas O-VAD produces grounded re- ports with localized objects, frame ranges, severity, and causal explanations. Why object state trajectories matter. Several patterns emerge. First, GPT- 5’s failures stem not from poor language reasoning but from impoverished evi- dence: without object-wise state trajectories, it receives caption-level inputs and judges only whether “the behavior matches expectations,” confirming that the bottleneck lies in fine-grained evidence construction. Second, O-VAD’s multi- object tracking enables cross-object diagnostic reasoning (e.g., correlating shaft bending with screw stripping in Case 2), which is unavailable to frame-level ap- proaches. Third, open-ended state-change events provide spatiotemporally grounded evidence anchoring every claim to specific objects and frame ranges, making re- ports directly actionable for root-cause diagnosis. O-VAD13 GPT-5 Answer: “No anomaly detected.” "report": "anomaly_detected": false,"overall_confidence": 0.55, "reason": "reasoning": "step1_observation": ”A clear plastic bottle with a white cap held by a metal clamp on a metallic arm; the arm rotates and inverts the bottle; the bottle remains securely gripped with no deformation; rotation stops and the system becomes static; no leakage is observed throughout.", "step2_expectation": "For an industrial manipulation/rotation task involving a plastic bottle and clamp, normal behavior is: stable grasp (no slip), smooth rotation/inversion, no crushing/deformation of the plastic bottle beyond elastic/temporary effects, and no liquid leakage. The clamp/arm should maintain alignment and control during motion and come to rest without dropping the object.", "step3_comparison": ”.........The described behavior matches expectations (secure grip, controlled rotation, no deformation, no leakage). Therefore, there is no evidence-supported anomaly to report.", "step4_causation": "Not applicable: no evidence of a deviation/failure is present in the provided tracking outputs, and the caption reports normal operation.", "step5_classification": "No anomalies detected per available evidence ......indicates normal manipulation without damage/leakage).", "step6_severity": "none. With no detected deviations or damage/leakage reported, there is no indicated impact on task completion, safety, or quality.” , "summary": "No anomaly is supported by the provided machine- readable evidence: both TRACKED OBJECTS and STATE CHANGES DETECTED are empty, so no object/frame-grounded deviations can be verified. The video describes controlled rotation/inversion of a clamped clear plastic bottle with no deformation or leakage, which is consistent with expected normal behavior.", O-VAD (Ours) Answer: Yes, plastic water bottle, loss_of_containment_leakage,60-80frames, causal analysis. "report": "anomaly_detected": true, "num_anomalies": 1, "overall_severity": "high", "overall_confidence": 0.82, "anomalies": [ "anomaly_type": "material_anomaly","anomaly_subtype": "loss_of_containment_leakage", "severity": "high", "description": "Closed plastic bottles release liquid during manipulation, indicating rupture/puncture or seal failure. Evidence includes obj_id=1 deformation+material_releaseat frames 60-70 (liquid pouring out) and continued material_releaseat 70-80 (bottle becomes almost empty).", "affected_objects": ["1”], "evidence_frames": [60,70,80], "start_frame": 60, "end_frame": 80, "confidence": 0.85, "reasoning_trace": "step1_observation": "Object obj_id=1 (Plastic water bottle; initial_state: intact, closed with cap; material: transparent plastic + water) shows repeated clamp- induced deformation across multiple intervals: 0-10, 10-20, 20-30, 30-40, 40-50, 50-60 (all deformation, moderate). A major event occurs at 60-70 where obj_id=1 has deformation + material_release(severe): the bottle is significantly compressed/bent and liquid is visibly pouring out. This continues at 70-80 as material_release(severe) with the bottle becoming almost empty......, At 150-160, a block presses the bottle and then moves away with the bottle returning to undeformed shape (moderate).", "step2_expectation": "Given the task context of industrial manipulation and the objects (thin-walled transparent plastic bottles containing water andclosed with caps), normal behavior would be: the clamp applies sufficient force to hold the bottle during rotation without causing significant permanent deformation, puncture, or leakage. Minor elastic indentation can occur with plastic, but it should not progress to severe crumpling or any material_release. A closed bottle should not leak during handling/rotation unless the cap is open/defective or the bottle is punctured/cracked. The gripping mechanism should also remain rigid (no bending/flexing) under normal operating loads.", "step3_comparison": "The observed behavior deviates from expectations. Leakage/material release: obj_id=1 shows severe material_releaseat 60-70 and 70- 80 (liquid pouring out; bottle becomes almost empty). Any material_releasefrom a supposedly closed, intact bottle during routine manipulation is abnormal and indicates loss of containment (cap failure, puncture, or crack).......", "step4_causation": "Plausible causes consistent with the tracked events: -Overclamping/ excessive grip force: Repeated deformation across many frame ranges (e.g., obj_id=1 from 0-60) suggests the clamp is applying too much pressure for a thin PET bottle, potentially leading to stress concentration and eventual rupture. -Misalignment or sharp/edge contact on the clamp: Localized dents and severe crumpling can occur if the clamp contacts the bottle with an edge or small area, increasing local stress and causing puncture. -Bottle integrity defect: A pre-existing weak seam, microcrack, or thin spot could fail under normal clamping, but the widespread severe deformation points more toward process/equipment force issues. -Cap/seal failure is less consistent with the evidence because multiple events describe leakage from the bottom (obj_id=2 at 60-70; obj_id=3 at 60-70) and pouring out (obj_id=1 at 60-70), which is more indicative of puncture/tear rather than cap loosening. -Manipulator/gripper mechanical issue: The reported bending/flexing of the gripping mechanism ..... which could also contribute to uncontrolled force application.", "step5_classification": "Anomalies detected: -material_anomaly: loss_of_containment/leakage from closed bottle(s) (obj_id=1 at 60-80). - deformation_anomaly: overcompression/structural damage risk due to excessive clamping (notably severe deformation at obj_id=1 60-70).....", "step6_severity": "Overall severity is high. The material_releaseevents (obj_id=1 severe at 60-80) indicate real product damage and process failure (spillage, contamination risk, cleanup/downtime). The repeated severe deformation suggests the process is not within safe handling limits and may continue to damage products. ......" "summary": "Tracking data indicates abnormal handling: The water bottle undergoes repeated clamp-induced deformation, including severe compression (e.g., obj_id=1 at 60-70). Critically, there is clear loss of containment with material_release: obj_id=1 pours out liquid at 60-70 and is nearly empty by 70-80. These deviations represent real product damage and process/equipment faults, yielding a high-severity anomaly." GPT-5 Answer: “No anomaly detected.” "report": "anomaly_detected": false, "overall_confidence": 0.62, "reason": "reasoning": "step1_observation": ” This video shows a screwdriver engaging and rotating a screw to tighten it into a metal hinge while the hinge remains static, ending with the screw fully tightened and motion stopping.", "step2_expectation": "For an industrial screwdriving/tightening operation on a metal hinge, normal behavior is: the driver aligns with the screw head, rotates smoothly, the screw advances downward into the hinge until seated, and then rotation stops (or torque-limits) without visible damage (e.g., no stripping, nThehinge should remain stationary if fixturedohinge deformation, no tool slip)..", "step3_comparison": ”........the behavior matches expectations (secure grip, controlled rotation, no deformation, no leakage). Therefore, there is no evidence-supported anomaly to report." "step4_causation": "Not applicable: no evidence of a deviation/failure is present in the provided tracking outputs, and the caption reports normal operation.", "step5_classification": "No anomalies detected per available evidence..... indicates normal manipulation without damage).", "step6_severity": "none. With no detected deviations or damage/leakage reported, there is no indicated impact on task completion, safety, or quality.” , "summary": "No anomaly is supported by the provided machine- readable evidence, so no object/frame-grounded deviations can be verified. The video caption describes controlled rotation/inversion of a clamped clear plastic bottle with no deformation or leakage, which is consistent with expected normal behavior.", O-VAD (Ours) Answer: Yes, metal hinge/screw, fastener_head_stripping,80-90frames, causal analysis. "report": "anomaly_detected": true, "num_anomalies": 3, "overall_severity": "high", "overall_confidence": 0.83, "anomalies": [ "anomaly_type": "deformation_anomaly","anomaly_subtype": "fastener_head_stripping", "severity": "high","description": "obj_id2 (Metal hinge / screw) shows clear screw-head slot stripping/rounding at frames 80-90, indicating irreversible damage likely from cam-out or excessive torque. This can prevent proper tightening and future removal, compromising assembly quality.", "affected_objects": [”2”], "evidence_frames": [80, 90], "confidence": 0.85, "reasoning_trace": "step1_observation": "Objects: obj_id1 (Door hinge, metal, intact) and obj_id2 (Metal hinge, metal, intact) remain the workpiece; obj_id3 is the screwdriver tip/shaft engaged with the screw (metal tip/screw, plastic handle). Multiple state changes are reported on the screw/driver interaction: (a) obj_id2 shows the screw rising upward (unscrewing/withdrawal) at frames 30-40 (deformation), again at 50-60 (deformation), and again at 100-110 (deformation). (b) obj_id2 shows explicit screw-head damage/stripping at frames 80-90(deformation: slot rounded/stripped). (c) obj_id3 shows the screw being driven in at 50-60, 70-80, 100- 110, 120-130, 150-160 (all logged as deformation but described as insertion/rotation), while also showing screwdriver shaft bending/flexing at 60-70 and 90-100 (deformation: shaft no longer straight). (d) obj_id3 shows the screw raised again at 170-179 (deformation: screw visibly raised/more exposed).The caption describes a normal tightening sequence ending with the screw fully tightened, but the tracked events include repeated reversals (up/down) and physical damage indicators.”, "step2_expectation": "For a metal hinge (obj_id1/2) being fastened with a screw and screwdriver (obj_id3), normal behavior is:the screw should monotonically advance downward into the hinge (decreasing exposed length) with smooth rotation until seated; the screwdriver shaft/bit should remain straight (no bending) under normal torque; the screw head/slot should remain intact (no stripping) if the correct bit is used, alignment is maintained, and torque is within spec. Minor visual changes from rotation are expected, but repeated unscrewing during a tightening operation and irreversible head damage are not expected.", "step3_comparison": "Deviations from expected fastening: (1) Direction reversals/withdrawal: obj_id2 reports the screw rising (unscrewing) at 30-40, 50-60, and 100-110, and obj_id3 reports the screw raised again at 170-179. In a standard tightening step, the screw should not repeatedly back out. (2) Irreversible fastener damage: obj_id2 at 80-90 reports stripping/rounding of the screw head slot, which is a real defect that can prevent achieving proper torque and future serviceability. (3) Tool deformation: obj_id3 at 60-70 and 90-100 reports the screwdriver shaft bending/flexing; a screwdriver shaft/bit should not visibly bend in normal operation, suggesting excessive torque, misalignment, or tool/bit failure. While some insertion events (obj_id3 at 70-80, 120-130, 150-160) match expectations, the combination of back-and-forth motion plus stripping and bending indicates abnormal process conditions rather than a clean tightening cycle.", "step4_causation": "Plausible causes include: (a) Bit-to-screw mismatch (wrong driver type/size) or worn bit causing cam-out, leading to stripping at 80-90 ....... (b) Misalignment between screwdriver (obj_id3) and screw axis causing side-loading; this can explain reported shaft ........ (c) Excessive torque or lack of torque control (no clutch/incorrect setting) causing both tool flex and screw head damage. (d) Cross-threading or obstruction in the hinge hole causing the operator/tool to reverse direction repeatedly to re-seat the screw....... (e) Workholdinginstability (hinge not firmly fixtured) could cause wobble and cam-out, ......", "step5_classification": "step5_classification": "Anomalies meeting the criteria (real damage/process failure): (1) deformation_anomaly: stripped screw head (fastener damage) on obj_id2 at 80-90. (2) manipulation_failure/ tool_integrity_anomaly: screwdriver shaft/bit bending under load on obj_id3 at 60-70 and 90- 100, .... (3) process_anomaly: repeated screw back-out/unscrewing during fastening......", "step6_severity": "Overall severity is high. The screw-head stripping at 80-90 is irreversible fastener damage with direct quality impact ......" "summary": "Tracking data indicates multiple abnormal events during hinge screwdriving: the screw repeatedly backs out (obj_id2 at 30-40, 50-60, 100-110; obj_id3 at 170-179), the screwdriver shaft bends under load (obj_id3 at 60-70 and 90-100), and the screw head becomes stripped/rounded (obj_id2 at 80-90). These deviations from a normal monotonic tightening process imply engagement/torque-control problems and result in irreversible fastener damage, creating a high-severity quality risk for the hinge assembly." Fig. 3: Qualitative results. Each row shows a different industrial manipulation task— plastic bottle rotation (top), hinge screw fastening (bottom)—with GPT-5’s output (left) and O-VAD’s output (right). Open-ended anomaly explanation. O-VAD can detect free-form anomalies grounded in tracked object states rather than a fixed taxonomy. In Case 1, it attributes the leakage to progressive stress from repeated clamp-induced defor- mation, citing the spatial relationship between gripper and breach. In Case 2, it synthesizes three independent anomaly streams (direction reversals, head strip- ping, shaft bending) into a coherent torque-control failure hypothesis. Such struc- tured causal reasoning goes beyond binary labels to provide the diagnostic detail operators need for corrective action. 14M. Yuan et al. 4.4 Ablation Studies We validate the contribution of each component in O-VAD by systematically removing one module at a time on four representative task subsets from Phys- AD (sticky roller, liquid, screw, and rubber band) that span diverse anomaly characteristics. Results are reported in Table 3. State Tracking is indispensable. Removing state tracking causes catas- trophic failure: precision, recall, and F1 drop to zero on three of four categories (sticky roller, liquid, screw) as the system collapses to predicting all samples normal, and recall falls to 0.033 (F1 0.065) on rubber band. Without an ex- plicit mechanism to model state transitions across frames, the system cannot distinguish anomalous deviations from normal progressions. BERT scores stay paradoxically high (0.867–0.947) as the model emits fluent but content-agnostic “normal” descriptions, underscoring that text-quality metrics alone are insuffi- cient for evaluating anomaly detection. Video-level captioning provides a global semantic prior. Removing video- level captioning has a category-dependent effect, revealing its role as a scene-level contextual anchor. For sticky roller and rubber band, where anomalies are holis- tic behavioral deviations (e.g., irregular rolling, band detachment), the caption supplies scene semantics that guide reasoning, and its removal causes AUC drops of 0.247 and 0.131. For liquid, classification metrics are unchanged and AUC de- creases marginally (0.034), so state tracking alone suffices for anomalies such as overflow. On screw, removing the caption slightly improves F1 (0.700→0.733): screw anomalies are localized and instantaneous (e.g., misalignment, incomplete tightening), so a coarse video-level summary introduces task-irrelevant context— motivating future multi-granularity captioning adapted to the anomaly’s spatio- temporal scale. Structured cognitive reasoning improves calibration. Replacing our cas- caded six-step reasoning module (Sec. 3.4) with a generic “think step by step” prompt still yields reasonable results, but the structured pipeline consistently im- proves AUC across all four categories (by 0.038–0.091) and raises BERT scores. Decomposing reasoning into process understanding, observation, expectation, comparison, causation, and classification thus yields better-calibrated confidence and more faithful explanations grounded in tracking evidence. Post-verification serves as an effective safeguard. Removing visual ver- ification (Sec. 3.4) preserves most hard classification metrics but consistently degrades AUC (up to 0.178 on sticky roller) and BERT scores, indicating it mainly benefits borderline cases rather than flipping high-confidence predictions. Through multiplicative confidence gating (Eq. 10), the verifier re-examines evi- dence frames against the caption to downweight spurious detections, sharpening the separation between true anomalies and ambiguous cases and improving ex- planation faithfulness. O-VAD15 Table 3: Ablation study on Phys-AD subset. Per-category Ac- c/P/R/F1/AUC/BERT scores across ablation variants. The result per category with full design is bold; orange marks degradation from the full model. sticky rollerliquid VariantAccPRF1 AUC BERTAccPRF1 AUC BERT O-VAD (full)0.689 0.691 0.967 0.806 0.811 0.689 0.667 0.667 1.000 0.800 0.776 0.666 (a) w/o Caption0.6440.659 0.9670.7840.5640.6450.667 0.667 1.000 0.8000.7420.650 (b) w/o State Track.0.3110.0000.0000.0000.484 0.9150.3330.0000.0000.0000.667 0.867 (c) w/o CoT Reas. 0.6670.667 1.0000.8000.7200.6480.667 0.667 1.000 0.8000.7380.632 (d) w/o Post-verifier0.6670.667 1.0000.8000.6330.6450.667 0.667 1.000 0.8000.7180.648 screwrubber band VariantAccPRF1 AUC BERTAccPRF1 AUC BERT O-VAD (full)0.600 0.700 0.700 0.700 0.704 0.784 0.633 0.611 0.733 0.667 0.721 0.784 (a) w/o Caption0.644 0.733 0.733 0.733 0.731 0.7880.5000.5000.6670.5710.5900.776 (b) w/o State Track.0.3330.0000.0000.0000.400 0.9360.517 1.0000.0330.0650.587 0.947 (c) w/o CoT Reas.0.644 0.733 0.733 0.7330.6640.7810.6170.595 0.7330.6570.6810.766 (d) w/o Post-verifier0.622 0.710 0.733 0.7210.6220.7580.5830.5710.6670.6150.714 0.799 5 Conclusion Without labor-intensive annotations, our agentic framework delivers state-of- the-art performance on any level anomaly detection, surpassing frontier VLMs on quantitative and qualitative evaluations. This demonstrates a new paradigm for industrial video anomaly detection that bridges the gap between low-level pattern recognition and high-level cognitive reasoning, paving the way for more reliable, explainable, and generalizable industrial inspection systems. Limitations. First, the multi-stage agentic pipeline incurs substantially higher latency than single-forward-pass VAD models, scaling with the number of tracked objects per video. Second, O-VAD relies on the VLM’s internalized common- sense rather than domain-specific expert priors, so its performance declines on anomalies defined by invisible physical properties or precise specifications that cannot be inferred from visual appearance alone—few-shot in-context learning with specification examples is a promising direction to bridge this gap. Third, static or perception-ambiguous defects (e.g., stuck buttons, degaussed magnets) that produce visually plausible behavior remain a fundamental boundary of perception-driven reasoning, motivating future integration of lightweight domain priors and process specification references. Acknowledgements This work was supported in part by U.S. NSF grants DBI-2238093, DBI-2422619, IIS-2211597, and MCB-2205148. Yizhou Zhao was supported in part by the SoftBank Group–ARM Fellowship. 16M. Yuan et al. References 1. Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., et al.: Qwen3-vl technical report. arXiv preprint arXiv:2511.21631 (2025) 2. Carion, N., Gustafson, L., Hu, Y.T., Debnath, S., Hu, R., Suris, D., Ryali, C., Alwala, K.V., Khedr, H., Huang, A., et al.: Sam 3: Segment anything with concepts. arXiv preprint arXiv:2511.16719 (2025) 3. Chen, Z., Li, J., Liang, J., Tan, L., Guo, Y., Lu, C., Li, Y.L.: Mˆ 3-vos: Multi-phase, multi-transition, and multi-scenery video object segmentation. In: Proceedings of the Computer Vision and Pattern Recognition Conference. p. 29193–29202 (2025) 4. Cheng, H.K., Oh, S.W., Price, B., Lee, J.Y., Schwing, A.: Putting the object back into video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. p. 3151–3161 (2024) 5. Dabouei, A., Shibu, J.P., Dalal, V., Cao, C., MacWilliams, A., Kangas, J., Xu, M.: Deep video anomaly detection in automated laboratory setting. Expert Systems with Applications 271, 126581 (2025) 6. Defard, T., Setkov, A., Loesch, A., Audigier, R.: Padim: a patch distribution model- ing framework for anomaly detection and localization. In: International conference on pattern recognition. p. 475–489. Springer (2021) 7. Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidi- rectional transformers for language understanding. In: Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers). p. 4171–4186 (2019) 8. Fan, L., Huang, J., Di, D., Su, A., Pagnucco, M., Song, Y.: Revitalizing reconstruc- tion models for multi-class anomaly detection via class-aware contrastive learning. arXiv preprint arXiv:2412.04769 (2024) 9. Fang, Z., Wang, X., Li, H., Liu, J., Hu, Q., Xiao, J.: Fastrecon: Few-shot in- dustrial anomaly detection via fast feature reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. p. 17481–17490 (2023) 10. Gu, Z., Zhu, B., Zhu, G., Chen, Y., Tang, M., Wang, J.: Anomalygpt: Detecting industrial anomalies using large vision-language models. In: Proceedings of the AAAI conference on artificial intelligence. vol. 38, p. 1932–1940 (2024) 11. Guo, J., Lu, S., Zhang, W., Chen, F., Li, H., Liao, H.: Dinomaly: The less is more philosophy in multi-class unsupervised anomaly detection. In: Proceedings of the Computer Vision and Pattern Recognition Conference. p. 20405–20415 (2025) 12. He, H., Bai, Y., Zhang, J., He, Q., Chen, H., Gan, Z., Wang, C., Li, X., Tian, G., Xie, L.: Mambaad: Exploring state space models for multi-class unsupervised anomaly detection. Advances in Neural Information Processing Systems 37, 71162– 71187 (2024) 13. Huang, C., Liu, Y., Zhang, Z., Liu, C., Wen, J., Xu, Y., Wang, Y.: Hierarchi- cal graph embedded pose regularity learning via spatio-temporal transformer for abnormal behavior detection. In: Proceedings of the 30th ACM international con- ference on multimedia. p. 307–315 (2022) 14. Huang, C., Wang, B., Wen, J., Liu, C., Wang, W., Shen, L., Cao, X.: Vad-r1: To- wards video anomaly reasoning via perception-to-cognition chain-of-thought. arXiv preprint arXiv:2505.19877 (2025) 15. Hyun, J., Kim, S., Jeon, G., Kim, S.H., Bae, K., Kang, B.J.: Reconpatch: Con- trastive patch representation learning for industrial anomaly detection. In: Pro- O-VAD17 ceedings of the IEEE/CVF winter conference on applications of computer vision. p. 2052–2061 (2024) 16. Jiang, X., Li, J., Deng, H., Liu, Y., Gao, B.B., Zhou, Y., Li, J., Wang, C., Zheng, F.: Mmad: A comprehensive benchmark for multimodal large language models in industrial anomaly detection. arXiv preprint arXiv:2410.09453 (2024) 17. Jiang, X., Liu, J., Wang, J., Nie, Q., Wu, K., Liu, Y., Wang, C., Zheng, F.: Soft- patch: Unsupervised anomaly detection with noisy data. Advances in Neural In- formation Processing Systems 35, 15433–15445 (2022) 18. Kang, H., Lee, W., Kim, J., Park, H.: Judo: A juxtaposed domain-oriented multi- modal reasoner for industrial anomaly qa. In: The Fourteenth International Con- ference on Learning Representations (2026) 19. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF international conference on computer vision. p. 4015–4026 (2023) 20. Li, G., Cai, G., Zeng, X., Zhao, R.: Scale-aware spatio-temporal relation learning for video anomaly detection. In: European Conference on Computer Vision. p. 333–350. Springer (2022) 21. Li, J., Dang, L., Xiao, Q., Shang, S., Cheng, J., Wu, H., Hao, Y., Wu, Q.: Video anomaly detection with semantics-aware information bottleneck. arXiv preprint arXiv:2506.02535 (2025) 22. Li, W., Xu, Y., Rao, Y., Wang, Z., Deng, S.: Vadtree: Explainable training-free video anomaly detection via hierarchical granularity-aware tree. Advances in Neu- ral Information Processing Systems 38, 148372–148404 (2026) 23. Li, W., Gu, Y., Chen, X., Xu, X., Hu, M., Huang, X., Wu, Y.: Towards visual discrimination and reasoning of real-world physical dynamics: Physics-grounded anomaly detection. In: Proceedings of the Computer Vision and Pattern Recogni- tion Conference. p. 30409–30419 (2025) 24. Lin, D., Qu, M., Han, K., Jiao, J., Jin, X., Wei, Y.: A unified reasoning frame- work for holistic zero-shot video anomaly analysis. arXiv preprint arXiv:2511.00962 (2025) 25. Lin, S., Wang, C., Ding, X., Wang, Y., Du, B., Song, L., Wang, C., Liu, H.: A vlm- based method for visual anomaly detection in robotic scientific laboratories. In: 2025 International Conference on Advanced Robotics and Mechatronics (ICARM). p. 34–39. IEEE (2025) 26. Liu, J., Yan, Y., Li, J., Zhao, W., Chu, P., Sheng, X., Liu, Y., Yang, X.: Ipad: Industrial process anomaly detection dataset. IEEE Transactions on Circuits and Systems for Video Technology 35(1), 380–393 (2024) 27. Liu, K., Ma, H.: Exploring background-bias for anomaly detection in surveillance videos. In: Proceedings of the 27th ACM International Conference on Multimedia. p. 1490–1499 (2019) 28. Park, H., Noh, J., Ham, B.: Learning memory-guided normality for anomaly detec- tion. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. p. 14372–14381 (2020) 29. Qi, L., Kuen, J., Guo, W., Shen, T., Gu, J., Jia, J., Lin, Z., Yang, M.H.: High- quality entity segmentation. arXiv preprint arXiv:2211.05776 (2022) 30. Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. p. 8748–8763. PmLR (2021) 18M. Yuan et al. 31. Ravi, N., Gabeur, V., Hu, Y.T., Hu, R., Ryali, C., Ma, T., Khedr, H., Rädle, R., Rolland, C., Gustafson, L., et al.: Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714 (2024) 32. Roth, K., Pemula, L., Zepeda, J., Schölkopf, B., Brox, T., Gehler, P.: Towards total recall in industrial anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. p. 14318–14328 (2022) 33. Singh, A., Fry, A., Perelman, A., Tart, A., Ganesh, A., El-Kishky, A., McLaughlin, A., Low, A., Ostrow, A., Ananthram, A., et al.: Openai gpt-5 system card. arXiv preprint arXiv:2601.03267 (2025) 34. Sun, S., Gong, X.: Long-short temporal co-teaching for weakly supervised video anomaly detection. In: 2023 IEEE International Conference on Multimedia and Expo (ICME). p. 2711–2716. IEEE (2023) 35. Sun, Y., Yang, X., Sun, J.J., Hariharan, B.: Tracking and understanding object transformations. arXiv preprint arXiv:2511.04678 (2025) 36. Tokmakov, P., Li, J., Gaidon, A.: Breaking the" object" in video object segmen- tation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. p. 22836–22845 (2023) 37. Wu, J.C., Hsieh, H.Y., Chen, D.J., Fuh, C.S., Liu, T.L.: Self-supervised sparse representation for video anomaly detection. In: European Conference on Computer Vision. p. 729–745. Springer (2022) 38. Wu, J., Zhang, W., Li, G., Wu, W., Tan, X., Li, Y., Ding, E., Lin, L.: Weakly- supervised spatio-temporal anomaly detection in surveillance video. arXiv preprint arXiv:2108.03825 (2021) 39. Wu, P., Zhou, X., Pang, G., Yang, Z., Yan, Q., Wang, P., Zhang, Y.: Weakly supervised video anomaly detection and localization with spatio-temporal prompts. In: Proceedings of the 32nd ACM International Conference on Multimedia. p. 9301–9310 (2024) 40. Ye, M., Liu, W., He, P.: Vera: Explainable video anomaly detection via verbalized learning of vision-language models. In: Proceedings of the Computer Vision and Pattern Recognition Conference. p. 8679–8688 (2025) 41. Yu, J., Li, X., Zhao, X., Zhang, H., Wang, Y.X.: Video state-changing object seg- mentation. In: Proceedings of the IEEE/CVF international conference on computer vision. p. 20439–20448 (2023) 42. Zavrtanik, V., Kristan, M., Skočaj, D.: Draem-a discriminatively trained re- construction embedding for surface anomaly detection. In: Proceedings of the IEEE/CVF international conference on computer vision. p. 8330–8339 (2021) 43. Zhang, H., Wang, Z., Wu, Z., Jiang, Y.: Diffusionad: norm-guided one-step denois- ing diffusion for anomaly detection (2023) 44. Zhang, X., Xu, M., Zhou, X.: Realnet: A feature selection network with realis- tic synthetic anomaly for anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. p. 16699–16708 (2024) 45. Zhao, S., Lin, Y., Han, L., Zhao, Y., Wei, Y.: Omniad: Detect and understand industrial anomaly via multimodal reasoning. arXiv preprint arXiv:2505.22039 (2025) 46. Zheng, H., Lin, T., Wang, W., Wang, Z., Zhang, W., Zhu, J., Shao, F.: Iad-unify: A region-grounded unified model for industrial anomaly segmentation, understand- ing, and generation. arXiv preprint arXiv:2604.12440 (2026) 47. Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al.: Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in neural information processing systems 36, 46595–46623 (2023) O-VAD19 48. Zhu, L., Chen, Q., Shen, X., Cun, X.: Vau-r1: Advancing video anomaly under- standing via reinforcement fine-tuning. arXiv preprint arXiv:2505.23504 (2025) 49. Zou, S., Tian, X., Wesemann, L., Waschkowski, F., Yang, Z., Zhang, J.: Unlocking vision-language models for video anomaly detection via fine-grained prompting. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. p. 4223–4233 (2026) 20M. Yuan et al. Appendix A More Details about Experiment Settings A.1 Details of Datasets. We evaluate O-VAD on Phys-AD [23] and LiquidAD [5], and additionally dis- cuss IPAD [26] to contextualize our design choices. All three target industrial video anomaly detection but differ in scale, interaction modality, and anomaly characteristics. Table 4 summarizes their key properties. Table 4: Comparison of industrial video anomaly detection datasets. DatasetDomainData Type #Videos #Obj. Categories #Anomaly Types FPS Annotation Granularity Periodicity Text Labels Phys-AD [23]Robot manipulationReal6,43422 (49 obj.)4760Video, Object✗✓ LiquidAD ‡ [5] Laboratory automationReal2,25118 (pipette-level)30Video, Frame, Object✗ IPAD [26]Factory equipmentSyn.+Real ∼2,000163925Video, Frame✓✗ ‡ Due to the difficulty of reproducing the frame- and pipette-level data splits of LiquidAD, we use only video-level annotations in this work. (i) Phys-AD [23] is a large-scale, physics-grounded video dataset for industrial anomaly detection. It is a real-world dataset focused on robotic arm manipu- lation, comprising 6,434 videos that cover 47 anomaly types across 22 object categories spanning metals, plastics, fluids, and articulated assemblies. Each ob- ject is paired with category-specific mechanical interactions, including press- ing, rotating, and stretching, performed by a UR5 robotic arm or servo motors. Ground-truth annotations are provided at both the video level and object level, covering normal and anomalous conditions. Detecting anomalies requires reason- ing about how an object’s physical state evolves under mechanical interaction, e.g. whether a rubber band develops micro-cracks or a screw backs out dur- ing tightening. The dataset further distinguishes persistent anomalies (visible throughout the entire sequence) from intermittent ones (emerging only at spe- cific temporal points), demanding fine-grained temporal understanding. These properties make Phys-AD an ideal testbed for evaluating O-VAD’s object-centric state tracking capability. (i) LiquidAD [5] targets anomaly detection in automated laboratory liquid transfer. It comprises 2,251 videos (1,792 training / 459 test) at 30 FPS, each capturing a complete transfer process in which 8 parallel pipettes dispense liq- uid into 8 test tubes. Anomalies include pipette clogging, incomplete dispensing, and volume deviations, with ground-truth annotations provided at the video, frame, and individual pipette levels. Although LiquidAD covers a single proce- dural scenario, it presents substantial intra-scene complexity: up to 8 simulta- neous dispensing events occur per frame, yet only one or two pipettes may be anomalous at any given time. Visual differences between normal and anoma- lous transfers are often subtle, such as a marginally reduced liquid column or a momentary dispensing hesitation, requiring temporally precise and spatially O-VAD21 localized detection. O-VAD addresses this challenge by constructing per-pipette state trajectories that identify which pipette deviates from expected behavior and at which frames the deviation occurs. (i) IPAD [26] is a video anomaly detection dataset designed for industrial man- ufacturing processes, comprising 597,979 frames across around 2,000 video clips. The dataset covers a diverse set of industrial devices, including conveyor belts, lifters, cutters, and grippers, encompassing both synthetic data (12 devices) and real on-site data (4 devices). A distinctive characteristic of IPAD is its explicit annotation of periodicity, a hallmark of industrial equipment in which actions repeat in fixed cycles. Anomaly types span appearance changes, positional devia- tions, motion irregularities, and logical errors, amounting to 39 distinct anomaly types in total. Importantly, environmental variations such as lighting changes and camera jitter are labeled as normal conditions, with the explicit intent of distinguishing genuine anomalies from benign scene perturbations. We evalu- ate O-VAD on IPAD (see §B.2), though we note that this dataset emphasizes periodic structure and cross-domain transfer rather than physics-grounded phys- ical interactions. Unlike Phys-AD [23] and LiquidAD [5], where anomalies are grounded in physical commonsense and mechanical interactions, IPAD defines anomalies primarily through deviations in object appearance (color, shape, size) and spatial location relative to reference normal samples. To adapt O-VAD to this reference-based evaluation protocol, we incorporate reference normal scenes as part of the object-level common sense representation, while continuing to exclude known anomaly types during evaluation (see §B.2 for details). Summary. These three datasets represent a progression of complementary chal- lenges: Phys-AD requires physics-grounded reasoning over diverse mechanical interactions, LiquidAD demands per-object temporal localization within clut- tered and visually homogeneous scenes, and IPAD emphasizes periodic tempo- ral structure with synthetic-to-real generalization. O-VAD’s three-stage pipeline, consisting of object discovery, state tracking, and chain-of-thought reasoning, is principally designed to address the first two settings, where anomalies are rooted in physical commonsense and object-level state evolution. Notably, the temporal reasoning capacity inherent in state tracking, which models how object states evolve across frames, generalizes naturally to periodic structures, making O- VAD’s training-free design readily extensible to the reference-based evaluation protocol of IPAD. A.2 Hyperparameter Details. O-VAD inherits and extends the threshold-based filtering mechanism of Tubelet- Graph [35]. Table 5 summarizes every threshold used across the three stages, together with its default value and the stage in which it operates. Below we discuss each threshold, its design rationale, and sensitivity. 22M. Yuan et al. VLM detection confidence. We set τ conf,1 0.1. During VLM-grounded object discovery (§3.2), SAM3 [2] segments objects identified by the VLM, and we retain detections with confidence above 0.1. This threshold is intentionally very permissive to maximize recall at the discovery stage: missing a task-relevant object early in the pipeline is irrecoverable, whereas false-positive objects can be pruned during downstream state tracking and anomaly reasoning. Table 5: Summary of threshold hyperparameters used in O-VAD. Symbol NameDefinitionValue Stage Origin τ conf,1 VLM detection confidence Confidence threshold for retaining VLM-grounded object detections during discovery0.11 Ours τ coverage Coverage thresholdThreshold for initiating new tubelets to avoid redundant spatial tracking0.25 2 [35] τ prox Spatial proximity gate Threshold ensuring new tracks emerge spatially near existing objects0.32 [35] τ sem Semantic consistency gate Threshold filtering tracks by CLIP semantic similarity to the target object0.72 [35] τ hi High-confidence bypass Threshold skipping verification for already overwhelmingly confident anomaly detections 0.83 Ours τ lo Low-confidence discard Threshold discarding low-confidence anomaly candidates as likely noise0.23 Ours τ conf Post-verification retention Threshold retaining verified anomalies after multiplicative confidence recalibration0.33 Ours Coverage threshold. We set τ coverage 0.25. When constructing the spatiotemporal partition (Eq.(4)), a new tubelet is initiated at frame t > 1 only when less than τ coverage of a newly detected entity’s area is already covered by existing tubelets. This value ensures completeness without spawning redundant tracks for regions that are already well-covered. It ensures that any region that is mostly untracked receives its own tubelet. In the original TubeletGraph work, this threshold was set without further tuning and shown to be robust across VOST [36], M 3 -VOS [3], and VSCOS [41], so we adopt the same value. Spatial proximity threshold. We choose τ prox = 0.3. A candidate track C emerg- ing at frame s is retained only if its initial mask overlaps with at least one of SAM2 [31]’s three multi-mask candidates at that frame by more than 30% (Eq. 5). The physical motivation is that a transformed or fragmented object (e.g. a piece breaking off, liquid being released) initially appears near the original ob- ject. The threshold is deliberately permissive because SAM2’s multi-mask output already expands the effective search region to account for segmentation ambi- guity during transformation. A grid search over τ prox ∈0.1, 0.2, 0.3, 0.4, 0.5 on the VOST train split shows that tracking performance (J ) varies by at most 1.5 points across this range [35], confirming low sensitivity. Semantic consistency threshold. We use τ sem = 0.7. A candidate track must exhibit a maximum masked CLIP [30] cosine similarity exceeding 0.7 with the prompt object across all valid frame pairs (Eq. 6). This threshold is stricter than τ prox because semantic filtering is the primary defence against false positives: spatially proximal but semantically unrelated entities (e.g. an operator’s hand entering the scene) must be rejected. At the same time, the threshold cannot be too aggressive (e.g. 0.9), because genuinely related post-transformation objects (e.g. foil box → foil sheet, intact bottle → deformed bottle with liquid) often O-VAD23 have only moderate CLIP similarity due to appearance change. The same grid search shows that τ sem = 0.7 is near-optimal and performance is stable across the [0.5, 0.8] range, dropping noticeably only at 0.9. High-confidence bypass. We use τ hi = 0.8. Anomalies with c orig > τ hi bypass visual verification entirely given the original confidence c orig . These represent cases where the accumulated state-change evidence is already overwhelming (e.g. clearly visible leakage or material release); re-verifying would add latency without altering the decision. Low-confidence discard. We use τ lo = 0.2. Anomalies with c orig < τ lo are dis- carded without verification. The reasoning chain itself is not confident in these candidates, and empirically they correspond to noise in the state-change events rather than genuine anomalies. Post-verification retention. We use τ conf = 0.3. For the intermediate band τ lo ≤ c orig ≤ τ hi , the final confidence after multiplicative gating exceeds 0.3 for the anomaly to be retained. The ablation study (Table 3) shows that removing the post-verifier degrades AUC by up to 0.178 (e.g.sticky roller), confirming that this recalibration sharpens the separation between true anomalies and ambiguous cases without flipping high-confidence predictions. A.3 Stage-wise Sampling and Grounding Details Frame sampling across stages. O-VAD applies a distinct, mostly deterministic sampling rule at each stage, summarized in Table 6. Stage 1 queries the VLM for the object inventory once on frame 0 and selects the SAM mask-prompt frame from 5 uniformly sampled candidates (step =N/5), retaining all frames for downstream use. Stage 2 subsamples at a content-adaptive frame rate set by an inter-stage VLM call, clamped to [2, 10] FPS (defaulting to 3 if the VLM is unavailable) to preserve key state transitions while reducing redundancy. Stage 3 loads frames at a fixed dataset-specific stride (60 / 30 / 25 for Phys-AD / LiquidAD / IPAD) for visual verification. All rules are deterministic except the Stage-2 rate, which is content-adaptive. Table 6: Frame sampling policy across O-VAD stages. StageSampling ruleRate / countDeterminism 1. Grounding Uniform scan + max-confidence selection 5 frames for SAM prompt; all frames retained deterministic 2. Tracking Inter-stage VLM, content-adaptive FPS FPS∈[2, 10] (default 3 if VLM unavailable)content-adaptive 3. Reasoning Dataset-specific stride for verification60 / 30 / 25 (Phys-AD / LiquidAD / IPAD)deterministic 24M. Yuan et al. Object deduplication (whole vs. parts). Stage 1 issues a single open-vocabulary VLM query requesting all distinct objects, and SAM is invoked with the full name list as a multi-class query that produces one fused binary mask. Whole and part labels (e.g., “screw” and “screw thread”) therefore collapse into a single connected mask region rather than spawning two tubelets. The textual labels are retained verbatim only as prompts for downstream per-tubelet identification, not as identity counts. Because naming is performed exactly once (on frame 0), no inter-frame label fusion is required in Stage 1; persistent identity is established later by SAM2 tracking in Stage 2. Spatial cues. Stage 1 persists two artifacts: a binary segmentation mask over the selected frame and an object-name list. Richer per-object spatial descrip- tors (bounding box, refined state, material) are produced downstream by the Stage 2 per-tubelet identification prompt, which returns structured trajectories with OBJECT/STATE/MATERIAL/BBOX fields. Spatial proximity and deformation. The Stage-2 proximity and semantic-consistency formulation is inherited from TubeletGraph [35]. Two properties make it robust under large deformation: (i) proximity is computed against SAM2’s three multi- mask candidates, not the single current-frame mask, explicitly accounting for segmentation ambiguity during transformation; and (i) semantic consistency uses CLIP mask-pooled features rather than raw appearance, tolerating visual change while preserving identity. A.4 Anomaly Type Evaluation. BERT similarity. Since O-VAD produces open-ended anomaly type labels, exact- match accuracy is inapplicable. We instead measure semantic alignment using BERT-base-uncased [7] cosine similarity. For each video, a reference string is the ground-truth anomaly-type label (or “normal” for non-anomalous videos), and a candidate string concatenates the predicted anomaly_type and anomaly_subtype fields across all detected anomalies (semicolon-delimited), defaulting to “normal” when none is detected. Both strings are encoded via the [CLS] embedding from BERT’s final hidden layer. The reported BERT score is the mean cosine simi- larity over all matched video–report pairs (covering both normal and abnormal ground truths). This metric naturally accommodates synonyms and paraphrases (e.g. “fastener head stripping” vs. “screw stripping”), and it should be read jointly with classification metrics. A.5 Full prompts of proprietary models. Prompts in Stage 1. During object grounding, the VLM is prompted to enumerate all task-relevant objects in each sampled frame and return their de- scriptions as a structured JSON array for downstream segmentation with SAM. O-VAD25 Prompt for p detect (Stage 1: Object Detection) "role": "user", "content": ( f"I t k " "Analyze this image and list ALL distinct objects/items visible." "For each object, provide a SHORT, SPECIFIC description that could be used to identify it. Focus on: - Main objects (not background elements) - Objects that could be tracked in a video - Use specific descriptions (e.g., ‘red apple’ not just ‘fruit’)" "Return ONLY a JSON array of object descriptions, nothing else." "Example format: [\"red apple being cut\", \"kitchen knife with black handle\", \"green cutting board\"]" ) Prompts in Stage 2. Before tracking state changes, the VLM is queried once per object on its first visible frame to establish a semantic baseline comprising its name, physical state, and material. Prompt for p id (Stage 2: Object Identification) "role": "user", "content": ( f" ̃ I i 1 " % highlighted first frame "Look at this image from a video. The object of interest is highlighted with a init_c_name contour." "Please describe:" "1. Object name (be specific)" "2. Current state (intact, squeezed, opened, etc.)" "3. Material type (plastic, metal, liquid inside, etc.)" "4. Bounding Box (spatial information)." "Answer format: OBJECT: [name] STATE: [current state] MATERIAL: [material] BBOX: [bbox]" ) The VLM compares color-highlighted visualizations of the same object across two time steps and reports any obvious state change with an open-ended type, description, and severity label. 26M. Yuan et al. Prompt for p state (Stage 2: State Change Detection) "role": "system", "content": "You are an expert video analysis assistant specialized in detecting object state changes and interactions. Your tasks include: 1. Identifying objects and their current states; 2. Detecting OBVIOUS change in object appearance, shape, material or moving state (slight changes should be ignored); 3. Recognizing interactions between objects; 4. Check for dynamics like material flow, deformation, or leaking." "role": "user", "content": ( f" ̃ I i t k " f" ̃ I i t k+1 " "Compare these two frames from a video." "The same object is highlighted with init_c_name contour in both." "First image: Earlier frame Second image: Later frame" "Carefully check for OBVIOUS changes:" "- DEFORMATION: Is the object’s shape different?" "- MATERIAL_RELEASE: Is anything coming out of the object?" "- SURFACE_CHANGE: Any cracks, tears, openings, damage?" "- TEXTURE_CHANGE: Has the surface appearance changed?" "- MOVEMENT_CHANGE: Has the object changed between static and moving?" "- INTERACTION_CHANGE: Are the objects interactively move together?" "Answer format: STATE_CHANGED: [yes/no] CHANGE_TYPE: [deformation/material_release/surface_change/ size_change/none/...] CHANGE_DESCRIPTION: [describe what changed in detail] CHANGE_SEVERITY: [none/slight/moderate/severe]" ) When multiple objects coexist in a frame, the VLM is additionally queried to characterize inter-object interactions, identifying the contact type, force direc- tion, and visible effect on the manipulated object. Prompt for p inter (Stage 2: Interaction Analysis) "role": "user", "content": ( f" ̃ I t " % both objects highlighted "Analyze the interaction between objects in this image. Object 1 (being manipulated): init_c_name contour Object 2 (manipulating): query_c_name contour" "Describe the interaction:" O-VAD27 "1. CONTACT_TYPE: How are they touching?" "2. FORCE_DIRECTION: Where is force being applied?" "3. VISIBLE_EFFECT: What effect is visible on Object 1?" "4. ACTION_VERB: What action is happening?" "Answer format: CONTACT_TYPE: [type] FORCE_DIRECTION: [direction] VISIBLE_EFFECT: [effect description] ACTION_VERB: [single verb describing the action]" ) Prompts in Stage 3. The core anomaly reasoning prompt supplies the VLM with all accumulated evidence—object metadata, state change events, video cap- tion, and sampled frames—and instructs it to follow a six-step chain-of-thought from observation through severity assessment, producing free-form anomaly clas- sifications. Prompt for p CoT (Stage 3: Anomaly Reasoning) "role": "system", "content": "You are an expert anomaly detection system for industrial processes. You have access to object tracking data including object metadata (descriptions, materials, initial states), video caption, and fine-grained state change events detected across video frames. ... Be thorough but avoid false positives. Consider physical plausibility and context. Always reference specific object IDs, frame ranges, and state change events in your reasoning." "role": "user", "content": ( f"sampled frames" "TASK CONTEXT: task_context VIDEO CAPTION: caption TRACKED OBJECTS: object_info STATE CHANGES DETECTED: state_changes" "Anomalies should cause or reflect real flaw in object," "real flaw in pipeline or real damage to object." "Follow this 6-step reasoning chain:" "STEP 1 - OBSERVATION: What specific changes and events" " occurred? Reference tracked objects by their IDs." "STEP 2 - EXPECTATION: What should have happened given" " the task context and object materials?" "STEP 3 - COMPARISON: How do observed state changes" 28M. Yuan et al. " differ from expectations?" "STEP 4 - CAUSATION: What could cause these deviations?" " Reason freely – do NOT limit to predefined categories." "STEP 5 - CLASSIFICATION: Classify anomalies using your" " own judgment. You are free to name the anomaly type." "STEP 6 - SEVERITY: Rate severity (none/low/medium/" " high/critical) based on impact, safety, reversibility." "Output your analysis as JSON: reasoning, anomalies, is_anomalous, overall_severity, summary" ) Each candidate anomaly with intermediate confidence is verified against evidence frames and the video caption, explicitly checking whether the claimed anomaly is a genuine failure or a normal process behavior. Prompt for p verify (Stage 3: Visual Verification) "role": "user", "content": ( f"evidence frames" "Video Caption: caption" "CLAIMED ANOMALY: Type: anomaly_type Description: anomaly_description Affected objects: affected_objects" "Based on the video caption and frames, verify if this" "CLAIMED ANOMALY really exists following these steps:" "1. Check Video Caption, do you find evidence of the" " claimed anomaly in caption or frames?" "2. Is the description accurate according to either" " caption or frames?" "3. What is your confidence level?" "Output JSON: verified: true/false, confidence: 0.0–1.0." ) A.6 Segmentation and Tracking Model Selection. O-VAD employs three distinct segmentation and tracking models—CropFormer, SAM2, and SAM3—at different pipeline stages. Each choice is motivated by the specific input-output requirements of that stage, as summarised in Table 7 and discussed below. O-VAD29 Table 7: Segmentation and tracking models used in O-VAD and the requirements that motivate each choice. ModelStageTaskKey Requirement SAM3 [2]1Object grounding Concept-aware: bridge VLM text output → pixel masks CropFormer [29] 2 Entity segmentation Exhaustive class-agnostic dense partition of every frame SAM2 [31]2 Tubelet propagation Memory-based temporal tracking with multi-mask output Stage 1: SAM3 for Concept-Aware Segmentation. Stage 1 (§3.2) requires converting VLM-generated object inventories, including names, natural-language descriptions, and spatial cues, into precise pixel-level masks. This is a semantic- to-mask task where the input is a textual concept, not a geometric prompt (point, box, or scribble). SAM3 (Segment Anything with Concepts) [2] extends the SAM family with concept-level understanding, enabling it to segment objects directly from semantic descriptions. Neither the original SAM [19] nor SAM2 [31] supports text-conditioned segmentation; using them would require an additional step to convert VLM descriptions into point or box prompts, introducing error and complexity. SAM3’s concept-aware capability thus provides the most direct bridge between the VLM’s structured output and the per-object masks M i 1 = SAM3(I 1 ,o i ) required by downstream stages. Stage 2: CropFormer for Entity Segmentation. The spatiotemporal par- titioning step (§3.3) requires an exhaustive, class-agnostic spatial partition of each frame, where every visually distinct region must be segmented, not just the target object. This is critical because the partition forms the search space for recovering post-transformation objects since any region that is not segmented can never be recovered as a candidate tubelet. CropFormer [29] is an entity segmentation model specifically designed for high-quality, class-agnostic segmentation of all “things” in an image. Its crop- based architecture handles objects across a wide range of scales, which is impor- tant in industrial settings where anomaly-relevant details (e.g. a small leak or a stripped screw head) can be very small relative to the frame. The ablations in [35] shows that replacing CropFormer with SAM2 automasks degrades J by 1.7 points, primarily due to SAM’s reduced reliability for small objects. Critically, CropFormer is a per-frame model with no temporal component. This is by design: each frame’s partition is computed independently so that newly appearing entities (post-transformation fragments, released material) are detected as soon as they become visible, without requiring any temporal prior that might suppress them. Stage 2: SAM2 for Tubelet Propagation. Once per-frame entities are de- tected by CropFormer, each entity is tracked forward in time via SAM2 [31] to 30M. Yuan et al. form tubelets (Eq. 4). SAM2 is designed for video-native object segmentation with a memory-attention mechanism that maintains temporal consistency. Two additional properties make it particularly well-suited for this role: (i)Multi-mask output. SAM2 produces three candidate masks per frame, capturing segmenta- tion ambiguity. O-VAD’s spatial proximity prior (τ prox , Eq. 5) explicitly exploits these candidate masks to estimate where post-transformation objects might ap- pear, effectively expanding the search region during state changes. (i)Object con- fidence scores. SAM2 outputs per-frame object score logits that indicate tracking reliability. These scores serve as an additional signal for detecting when a track is being lost, which is a potential indicator of an ongoing state transformation. Meanwhile, the ablation in [35] also shows that replacing SAM2 with Cutie [4] for tubelet propagation causes a more significant degradation (3.3 in J and 9.3 in temporal recall T R ), confirming that SAM2’s tracking quality is critical for the partition’s completeness. Why Not a Single Model? Each stage has fundamentally different input- output requirements: Stage 1 needs text → mask (concept-aware segmentation), Stage 2’s partitioning needs frame→ all masks (exhaustive entity segmentation), and Stage 2’s tracking needs mask→ mask sequence (temporal propagation with memory). No single existing model satisfies all three. SAM3 can segment named objects but is not designed for exhaustive spatial partitioning; CropFormer pro- duces dense per-frame segments but does not perform temporal tracking; and SAM2 excels at temporal propagation from a given mask but requires geometric prompts and does not perform class-agnostic partitioning. The modular design allows each component to operate at its respective strength, and the clean in- terfaces between stages (masks, tubelets, state-change events) make the pipeline straightforward to update as better models become available. B Detailed Per-Category Quantitative Results B.1 Detailed Results on Phys-AD O-VAD matches or surpasses trained methods while dominating all training-free baselines across 22 categories. O-VAD achieves the highest average type-level BERT score (0.803) among all training-free methods and claims the best score on 14 of 22 categories, confirming that object-centric state tracking yields anomaly descriptions semantically closer to ground-truth labels. At the video level, O- VAD attains the best or second-best AUROC on 16 of 22 categories and even surpasses both trained baselines on five categories whose anomalies involve pro- gressive, multi-step state changes invisible to single-pass detection. The most pronounced advantage lies in recall. O-VAD nearly doubles Qwen3-VL-32B’s average and exceeds URF-ZS-HVAA’s by an order of magnitude, particularly on interaction-heavy categories where the state tracking module prevents the VLM from defaulting to “normal.” Precision remains comparable to Qwen3-VL-32B, though achieved through a fundamentally different mechanism. O-VAD detects aggressively and relies on post-verification to suppress false positives, rather O-VAD31 than abstaining under uncertainty. Overall, without any training data, domain knowledge, or predefined taxonomy, O-VAD delivers performance competitive with and sometimes exceeding fully supervised methods, while consistently out- performing all training-free alternatives. Trained baselines show uneven category coverage. While S3R [37] achieves the highest average AUROC (0.612) among all methods thanks to strong perfor- mance on categories with distinctive anomaly distributions (e.g., Roll. Bear. 0.963, Sticky Roller 0.864, Rub. Band 0.853), it collapses on categories requiring temporal reasoning (e.g., Hinge 0.108, Lock 0.361, Clip 0.587). MNAD.p [28] shows the complementary pattern: strong on temporal categories (Fan 0.881, Clock 0.763) but near-chance on others (Rub. Band 0.241, Screw 0.387). This category-dependent brittleness of trained methods, together with their inability to produce semantic rationale, underscores the practical advantage of O-VAD’s training-free, object-centric paradigm. Challenging categories and failure modes. O-VAD’s weakest video-level AUROC values occur on Caster Wheel (0.301), Zipper (0.388), Button (0.393), and Slide (0.414). These categories share characteristics that challenge perception-driven reasoning. Caster Wheel anomalies are often defined by subtle rotational resis- tance changes that produce minimal visual cues; Zipper and Button anomalies involve small, fast state transitions (e.g., a stuck zipper tooth, a button failing to click) that may be missed at the default 10-frame sampling interval; and Slide anomalies depend on precise spatial alignment that is difficult to assess from cropped object views. These failure modes motivate future work on adaptive temporal sampling and multi-scale spatial reasoning. More detailed analysis of failure cases can be found in the §C.4. B.2 Detailed Results on IPAD Given that IPAD is a comprehensive dataset for video anomaly detection in in- dustrial manufacturing, this subsection provides an extensive comparative anal- ysis to demonstrate the effectiveness of our O-VAD framework on this dataset. Table 9 reports video-level and frame-level AUROC, different training-free meth- ods across 16 industrial scenarios. Considering the special conditions discussed in §A.1, we modify the CoT prompt for anomaly reasoning to align O-VAD with IPAD benchmarking. Specifically, when prompting the VLM for anomaly reasoning, we inject one normal reference case (three evenly sampled frames) as additional model input and replace the "expectation" part of prompt with pointing to the reference frames accordingly. Such modification is intended to bridge the gap between the dataset’s anomaly types are grounded on comparison to normal cases and physical commonsense, which achieves the goal of incorpo- rating such comparison grounding sense without domain specific abnormal types specification. 32M. Yuan et al. O-VAD matches or surpasses all training-free baselines. O-VAD achieves the highest average video-level AUROC (56.5%) and similar frame-level AUROC (51.8%) among all training-free methods, which demonstrates O-VAD’s improve- ment upon simple querying on VLMs. More importantly, O-VAD also ranks the highest in average video-level Accuracy, Revall, F1 score and average frame-level Recall, F1 score. Specifically, O-VAD secures the best video-level performance among all methods in 12 out of 16 scenarios and frame-level in 8 out of 16 scenarios, which shows its high effectiveness in complex industrial scenarios. Challenging subsets and failure modes. O-VAD’s weakest video-level AUROC values occur on S05 (39.6%), S08 (38.9%), R02 (43.3%) and R03 (46.7%). These subsets share characteristics that challenge perception-driven reasoning. S05 anomalies are often defined by subtle changes in object location and orienta- tion, which can only be inferred from minimal visual cues. S08 are complex scenes containing multiple main objects to keep track of and requires rigorous reasoning on the underlying normal rules, which make object change detections not detailed enough. R02 and R03 subsets always have scenarios with multi- ple objects, making O-VAD hard to decide which objects should be treated as background and should not be considered as anomalies. C More Qualitative Results. C.1 Reliability of Intermediate Outputs State trajectory extraction. Lacking ground-truth state-change annotations, we assess Stage-2 reliability along three label-free axes. Count: Table 10 reports per-dataset source properties (duration, FPS) alongside per-stage intermedi- ate counts (Stage-1 grounded objects, Stage-2 detected states), summarized as min./avg./max. over the evaluation set. Grounded object counts track ground- truth cardinality (e.g., up to 8 pipettes on LiquidAD, a single foreground ob- ject on most Phys-AD categories) and per-video state counts exhibit bounded variance, ruling out both under-firing (collapse to “no change”) and over-firing (spurious-event flooding). Content: the human study (§C.5) confirms that cited state changes match the underlying video evidence. Pattern: on IPAD’s periodic normal cycles, detected events cluster at cycle boundaries rather than distribut- ing uniformly, providing a label-free signal of temporal fidelity. The “w/o State Tracking” ablation (Table 3) supplies a fourth axis via downstream impact, where recall collapses once the signal is removed. Post-verification threshold sensitivity. (τ hi ,τ lo ,τ conf ), the post-verifier’s three thresholds, encode a principled skip/verify/discard partition of the confidence space rather than three independent free parameters, leaving one effective hy- perparameter (the verification band). The “w/o Post-verifier” ablation (Table 3) already shows AUROC drops up to 0.178 on sticky roller, confirming its role in recalibrating borderline cases; a ±0.1 sensitivity sweep over each threshold leaves the retained anomaly set stable, indicating the pipeline is not sensitive to the exact cut points. O-VAD33 C.2 Multi-Instance Re-Identification Production scenes with many look-alike instances (e.g., LiquidAD’s 8 identical pipettes) raise two distinct sub-problems, each handled by a separate mechanism. Enumeration (Stage 1): we extend the open-vocabulary detector with an itera- tive scheme—the VLM emits a class-level bounding box, the highest-confidence instance inside is segmented and its pixels are masked out, and detection is re- invoked on the masked image until no further instance is returned—yielding one mask per look-alike instance by construction. Persistent identity (Stage 2): each enumerated instance is propagated by SAM2’s mask-memory attention, which maintains identity positionally rather than through appearance descriptors that would collapse on near-identical objects. Table 10 confirms that grounded counts match ground-truth cardinality on LiquidAD (8 pipettes per video), and Fig- ure 4 illustrates both stages under partial occlusion by the dispensing head and during liquid transfer. Fig. 4: Multi-instance re-ID on LiquidAD. O-VAD’s iterative detect–segment– mask-out scheme assigns one stable tubelet per pipette across all 8 visually-identical instances, including under partial occlusion by the dispensing head and during liquid transfer. C.3 Comparative Reasoning Traces We compare the full reasoning traces of O-VAD against baseline VLMs on two representative success cases from LiquidAD and Phys-AD (Fig. 5). LiquidAD: unintended single-channel drip. In this case, a multi-channel pipette head releases a droplet from only one tip at frames 30–40, while the remaining tips stay unchanged. Qwen3-VL-32B processes the full video and reports normal operation, noting that the movement is smooth, tips are properly seated, and 34M. Yuan et al. no spills or bubbles are observed. Its frame-level descriptions confirm correct aspiration and dispensing but miss the isolated single-tip drip entirely. This is because the anomaly is localized to one channel among many, making it invis- ible to holistic frame-level inspection. O-VAD, by contrast, tracks the pipette head as a single object with per-frame state annotations. At frames 30–40, the state tracker flags a material release event from only one tip (the second from the right). The reasoning chain then compares this against the expected multi- channel dispense pattern and concludes that an isolated drip is inconsistent with a controlled dispense step. It attributes the cause to a possible seal or pressure imbalance in that channel. This case demonstrates how object-centric tracking surfaces fine-grained per-component deviations that frame-level captioning over- looks. Phys-AD: sticky roller crushing. Here, a robotic gripper manipulates a paper towel roll mounted on a cardboard core. GPT-5 observes the roller rotating and stopping, describes the behavior as consistent with normal controlled manipula- tion, and reports no anomaly. Its reasoning explicitly states that no deviations can be established against expectations. O-VAD tracks three objects indepen- dently: the cardboard core (obj_id 1), a metal gripper component (obj_id 2), and the paper towel roll (obj_id 3). It detects repeated crushing and flattening of the cardboard core at frames 0–10, 60–70, 140–150, and 160–170, along with compression and bending of the paper roll at frames 110–140. It further identifies a surface tear on the roll end at frames 60–70 and 160–170. The reasoning chain synthesizes these multi-object observations into a coherent diagnosis: excessive grip force or misalignment is causing progressive product damage, supported by concurrent deformation flags on the gripper itself (frames 100–150). GPT-5 misses all of these because its caption-level analysis describes only the high-level motion (rotation and stop) without examining per-object physical integrity. Both cases illustrate the same bottleneck identified in the main paper: base- line VLMs generate fluent and plausible reasoning but lack fine-grained object- level evidence to anchor their judgments. O-VAD’s structured state trajectories provide the missing evidential layer, enabling the reasoning chain to detect sub- tle, localized anomalies and to correlate changes across multiple objects for causal diagnosis. C.4 Failure Cases Analysis We present four representative failure cases that reveal the fundamental limita- tions of perception-driven anomaly reasoning without predefined expert knowl- edge. These cases fall into two categories: specification-dependent anomalies, where the ground truth is defined by quantitative thresholds invisible to vi- sual inspection, and perception-ambiguous anomalies, where the defect produces visually plausible behavior indistinguishable from normality. Specification-dependent failures. In the servo angle-restricted case (Fig. 6, top), the servo motor rotates to a small angle and returns smoothly. Both Qwen3-VL- 32B and O-VAD classify this as normal, since the observed motion is visually O-VAD35 consistent with a correctly functioning servo. The anomaly is a restricted rotation range, which is defined by a quantitative angular specification entirely absent from the visual evidence. Without knowing the expected rotation angle, neither state tracking nor commonsense reasoning can flag this deviation. Similarly, in the clip unable-to-press case (Fig. 7, bottom), the pressing tool descends and retracts while the clip remains static. O-VAD detects no state change events for the clip and concludes no anomaly is present. The ground truth is that the clip should have been pressed but was not. This requires knowledge of the expected actuation outcome, which is not visually inferrable. Perception-ambiguous failures. In the button stuck case (Fig. 7, top), O-VAD tracks repeated deformation of the metallic pressing tool across multiple frame intervals. It flags this as tool/fixture damage due to excessive force. However, the actual anomaly is the button’s failure to actuate: the pressing tool deforms precisely because the button is stuck. Both a stuck button and a normally stiff button produce similar visual signatures of tool compression. O-VAD captures the correct symptom (tool deformation) but misattributes the cause, illustrating how accurate state tracking can still lead to incorrect classification when the root cause requires understanding the intended function of the pressed object. In the magnet degaussing case (Fig. 6, bottom), a degaussed magnet de- taches from the board during manipulation. Qwen3-VL-32B attributes this to a gripping mechanism failure. O-VAD detects over-clamping deformation and grip instability, identifying the anomaly as crushing and dropping rather than loss of magnetism. Since the magnet visually resembles an ordinary red plastic block, neither method can infer its magnetic properties. The detachment is caused by insufficient magnetic adhesion, not mechanical failure, but this distinction is invisible to visual perception. These failure cases share a common pattern: the ground-truth anomaly is defined by an invisible physical property (magnetic strength, required actuation force, angular specification) or requires functional knowledge of what the object is supposed to do. Such information cannot be recovered from visual appearance alone. While O-VAD’s object-centric state tracking correctly captures observable symptoms in most cases, it cannot bridge the gap between perceptual evidence and specification-level judgments. This motivates future work on integrating few-shot specification examples or lightweight expert priors into the reasoning pipeline. C.5 Human Evaluation Human Evaluation Protocol. To assess the practical utility of anomaly reports beyond automated metrics, we conduct a human evaluation study with N=5 domain experts (graduate researchers with industrial vision or robotics experi- ence). Each evaluator independently reviews anomaly reports generated by three methods—Qwen3-VL-32B (direct prompting), URF-ZS-HVAA, and O-VAD—on 10 randomly sampled videos from the Phys-AD test set, spanning 10 object cat- egories. Reports are anonymized and presented in randomized order to prevent 36M. Yuan et al. method-identification bias. Each evaluator completes the questionnaire below for every report. Instructions. Each evaluator should independently review anomaly reports gen- erated by three methods on 10 randomly sampled videos. For each video–report pair, first watch the video, then read reports, and rate each of the following on a 5-point Likert scale: 1 (strongly disagree) to 5 (strongly agree). Evaluation Questionnaire. The questionnaire (Table 11) comprises 10 ques- tions organized into five parts that progressively assess report quality from fac- tual correctness to practical utility. Part I (Q1–Q4) measures detection cor- rectness at four granularities: binary anomaly/normal classification, semantic anomaly type matching, affected object identification, and temporal localiza- tion. Part I (Q5–Q7) evaluates explanation quality along three complemen- tary axes: faithfulness (no hallucinations), completeness (no omissions), and causal coherence (logically sound reasoning). Part I (Q8) assesses actionability, i.e., whether the report alone suffices for an operator to diagnose and respond. Part IV (Q9) captures a holistic quality judgment. Finally, Part V (Q10) elic- its a direct pairwise preference ranking across all methods for the same video, yielding a deployment-oriented win rate. All dimensions except Q10 use a 1–5 Likert scale; Q10 produces a rank ordering from best to worst. Human Evaluation Results We report results from the human evaluation study described in Appendix C.5. Five domain experts independently rated anomaly reports from three methods—URF-ZS-HVAA (Report 1), Qwen3-VL- 32B (Report 2), and O-VAD (Report 3)—on 10 Phys-AD videos (7 anomalous, 3 normal) across 9 Likert-scale dimensions (Q1–Q9) and a pairwise preference ranking (Q10). Table 12 summarizes the results. Overall comparison. O-VAD achieves the highest scores on all nine dimensions and is ranked first by evaluators in 82% of cases (41 out of 50 evaluator–video pairs). Its average detection score (4.42) exceeds Qwen3-VL-32B [1] (2.89) by +1.53 points and URF-ZS-HVAA [24] (1.14) by +3.28 points. The gap is simi- larly pronounced for explanation quality (+1.11 over Qwen3, +3.32 over URF- ZS) and actionability (+1.32 over Qwen3, +3.16 over URF-ZS). URF-ZS-HVAA consistently scores near the floor (1.0–1.2) across all dimensions, confirming that its chained reasoning framework, despite incorporating temporal detection and spatial localization modules, produces reports that evaluators judge as almost entirely uninformative for industrial anomaly diagnosis. This is consistent with its near-zero recall in the quantitative evaluation (Table 1 and Table 2). Detection correctness. The most striking result is on binary detection (Q1), where O-VAD scores 4.98 out of 5—near-perfect agreement that it correctly iden- tifies anomalous vs. normal videos. In contrast, URF-ZS-HVAA scores only 1.24, indicating that evaluators find its predictions almost always incorrect. Qwen3- VL-32B scores 2.70, performing better than URF-ZS but still far below O-VAD. O-VAD37 O-VAD also leads on anomaly type identification (Q2: 4.26 vs. 2.60 vs. 1.08) and temporal localization (Q4: 3.96 vs. 2.32 vs. 1.08), though these finer-grained dimensions show more variance, reflecting the inherent difficulty of open-ended type classification and frame-level precision. Explanation quality and actionability. Among the explanation dimensions, O- VAD’s largest advantage is on completeness (Q6: 4.58 vs. 3.36 vs. 1.00), indi- cating that evaluators find its state-trajectory-grounded reports cover all rel- evant state changes without significant omissions. Causal coherence (Q7: 4.30 vs. 3.02 vs. 1.00) shows the second-largest gap, confirming that the structured chain-of-thought reasoning produces more logically sound root-cause analyses. On diagnostic usefulness (Q8), O-VAD scores 4.16, meaning evaluators believe an operator could largely act on the report without re-watching the video, while URF-ZS-HVAA’s score of 1.00 indicates its reports provide no actionable infor- mation. Anomalous vs. normal cases. On the 7 anomalous videos, O-VAD’s detection average (4.40) substantially exceeds Qwen3-VL-32B (3.19) and URF-ZS-HVAA (1.00). On the 3 normal videos, O-VAD achieves a perfect binary correctness score of 5.00, whereas both baselines score below 1.80 on Q1, indicating severe false-positive tendencies. Notably, Qwen3-VL-32B’s binary correctness drops to 1.13 on normal cases (compared to 3.37 on anomalous ones), suggesting it de- faults to predicting anomalies. O-VAD’s robustness on normal videos stems from its evidence-grounded reasoning: without tracked state changes that violate pro- cess expectations, the chain-of-thought correctly concludes normality. C.6 LLM-as-Judge LLM-as-Judge Evaluation Protocol. To enable scalable evaluation beyond the 10-video human study, we design an LLM-as-judge [47] protocol that mirrors the human questionnaire (§C.5) and can be applied to the full test set. We use GPT-4o as the judge model with structured prompts. LLM-as-Judge Results. To complement the human evaluation and scale as- sessment to a larger sample, we apply the LLM-as-judge protocol using GPT-4o as the judge model across the same 10 Phys-AD videos. Note that Q4 (tempo- ral localization) is excluded from this evaluation due to the lack of frame-level ground-truth annotations in the judging set. Table 12 reports the results. Overall comparison. O-VAD is ranked first by the LLM judge in 70% of cases, fol- lowed by Qwen3-VL-32B at 30% and URF-ZS-HVAA at 10%. O-VAD achieves the highest overall quality (Q9: 3.50 vs. 2.70 vs. 1.20) and diagnostic useful- ness (Q8: 3.40 vs. 2.30 vs. 1.10), consistent with the human evaluation trends. URF-ZS-HVAA again scores near the floor across all dimensions (1.00–1.40), confirming its inability to produce meaningful anomaly reports regardless of the evaluation paradigm. 38M. Yuan et al. Detection vs. explanation: a revealing split. The most notable finding is the di- vergence between detection correctness and explanation quality. O-VAD leads convincingly on detection (avg. 3.20 vs. 2.33 vs. 1.13) and binary correctness in particular (Q1: 4.60 vs. 2.60 vs. 1.40). However, Qwen3-VL-32B achieves higher explanation scores than O-VAD on all three dimensions: faithfulness (4.90 vs. 3.90), completeness (4.00 vs. 3.90), and causal coherence (3.70 vs. 3.60). This reversal, absent in the human evaluation where O-VAD leads on all explana- tion metrics, reflects a known LLM-judge bias toward fluent, well-structured prose [47]. Qwen3-VL-32B generates polished natural-language narratives even when its detection is incorrect, which the LLM judge rewards on explanation dimensions evaluated independently of correctness. Human experts, by contrast, penalize explanations that describe the wrong conclusion, regardless of fluency. Despite this explanation advantage, Qwen3-VL-32B’s poor detection correctness (Q1: 2.60) and low actionability (Q8: 2.30) result in a lower overall quality score and win rate, indicating that the LLM judge’s holistic assessment still correctly prioritizes factual accuracy. Anomaly type correctness remains challenging. Both O-VAD and Qwen3-VL- 32B tie on type correctness (Q2: 1.80), the lowest-scoring dimension for both methods. This reflects the inherent difficulty of open-ended anomaly type clas- sification: even when detection is correct, precisely matching the ground-truth anomaly taxonomy remains challenging for all methods, a gap that motivates future work on domain-adaptive type reasoning. Consistency with Human Evaluation. The LLM-as-judge results broadly corroborate the human evaluation findings: O-VAD ranks first overall, URF- ZS-HVAA is consistently the weakest, and O-VAD’s primary advantage lies in detection correctness and actionability. The main discrepancy is on explanation quality, where the LLM judge favors Qwen3-VL-32B’s fluent but often factually incorrect narratives. This highlights a complementary strength of human evalu- ation: experts naturally integrate correctness into their explanation judgments, producing assessments more aligned with real deployment requirements. D Algorithm of O-VAD Algorithm 1 summarizes the complete O-VAD pipeline. Stage 1 discovers objects via VLM prompting and segments them with SAM, Stage 2 tracks each object through transformations and queries the VLM on sampled frame pairs to de- tect open-ended state changes, and Stage 3 performs a six-step chain-of-thought reasoning pass followed by confidence-tiered visual verification to produce the final anomaly report. All three stages are training-free and require no predefined taxonomy. O-VAD39 Algorithm 1 O-VAD: Object-Centric Video Anomaly Detection Require: Video V =I t T t=1 , optional caption cap, optional task context ctx Ensure: Anomaly report R 1: 2: // Stage 1: Automated Object Grounding 3: Sample candidate frames I t k K k=1 from V 4: for each candidate frame I t k do 5: O t k ← φ VLM (I t k )▷ object inventory: names, descriptions, spatial cues 6: end for 7: O ← S k O t k ▷ merge and deduplicate 8: for each object o i ∈O do 9: M i 1 ← SAM(I 1 ,o i )▷ grounded segmentation 10: end for 11: 12: // Stage 2: Object-Centric State Tracking 13: // Spatiotemporal partitioning & tubelet construction [35] 14: P ← TubeletGraph(V, M 1 )▷ partition video into tubelets 15: // State tracking: recover missing tracks 16: for each object o i with prompt track P i do 17: T i ← P i ∪C∈P | S prox (C,P i )>τ prox ∧ S sem (C,P i )>τ sem 18: end for 19: // Initial state identification 20: for each object o i with track T i do 21: ̃ I i 1 ← Highlight(I 1 ,M i 1 ) 22: (desc i , state i , mat i )← φ VLM ( ̃ I i 1 )▷ object name, state, material 23: end for 24: // State change detection and understanding (ours) 25: E ←∅ 26: for each object o i with track T i do 27: for each sampled pair (t k , t k+1 ) from T i do 28: ̃ I t k ← Highlight(I t k ,M i t k ); ̃ I t k+1 ← Highlight(I t k+1 ,M i t k+1 ) 29:(d, ty, ca, desc, sev)← φ VLM ( ̃ I t k , ̃ I t k+1 )▷ open-ended state change query 30:if d = yes then 31:E ←E ∪(t k , t k+1 , ty, ca, desc, sev, i) 32:end if 33: end for 34: end for 35: 36: // Stage 3: State-Aware Anomaly Reasoning 37: 38: // Cognitive anomaly reasoning (single VLM call, Steps 1–6) 39: Sample key frames F =I 1 , I ⌊T/2⌋ , I T from V 40: (reasoning, A raw )← φ VLM O, E, ctx, cap, F ▷ Steps 1–6: observe → expect → compare → cause → classify → severity 41: 42: // Post visual verification 43: for each anomaly a∈A raw do 44: if c orig (a)≥ τ hi then 45:Retain a▷ high-confidence: skip verification 46: else if c orig (a)≤ τ lo then 47:Discard a▷ low-confidence: auto-discard 48: else 49:(verified, c ver )← φ VLM (F, cap, a)▷ check anomaly evidence in caption/frames 50:if verified = true then 51:c final (a)← c orig (a)· c ver 52:else 53:c final (a)← c orig (a)· (1− c ver ) 54:end if 55:if c final (a) < τ conf then 56:Discard a▷ low post-verification confidence 57:end if 58: end if 59: end for 60: A← retained anomalies 61: 62: return Report R = (A, E, reasoning, summary) 40M. Yuan et al. Table 8: Video- and type-level results on Phys-AD [23] across 22 object cate- gories. “†” denotes methods requiring training data; “∗” denotes training-free methods. “BERT” is the type-level BERTScore, reported only for methods that emit free-form anomaly-type descriptions. For each row, the best / 2nd-best across all methods is high- lighted: video-level metrics in /, type-level BERT in/; our best results are additionally in bold. Blank cells indicate results not available. The 22 categories are split into two side-by-side blocks; “Avg.” reports the average over all categories. Trad.DirectWkf.OursTrad.DirectWkf.Ours Cat.Met. MNAD † S3R † Qwen3 ∗ GPT5 ∗ URF ∗ VERA † O-VAD ∗ Cat.Met. MNAD † S3R † Qwen3 ∗ GPT5 ∗ URF ∗ VERA † O-VAD ∗ Ball Acc0.637 0.326 0.504 0.6770.704 0.4770.570 Magnet Acc0.533 0.767 0.5440.908 0.589 0.6310.689 P0.681 0.0000.695 0.667 0.692 0.6670.742P0.680 0.800 0.7710.968 0.7171.0000.700 R0.856 0.000 0.456 0.8001.000 0.0570.544R0.5670.867 0.450 0.857 0.633 0.3140.933 F1 0.759 0.000 0.550 0.7270.818 0.1050.628F10.6180.832 0.5680.909 0.673 0.4780.800 AUC0.589 0.155 0.5280.744 0.461 0.5480.579AUC 0.5120.724 0.592 0.364 0.5180.8920.576 BERT 0.8210.6420.862BERT0.7810.6430.699 Button Acc0.697 0.203 0.330 0.4000.580 0.4620.380 Roll. Bear. Acc0.433 0.483 0.4290.625 0.5170.5360.536 P0.8071.000 0.809 0.3750.967 0.5000.781P0.462 0.000 0.3500.8570.509 0.0000.500 R0.817 0.004 0.212 0.1710.492 0.0290.312R0.800 0.000 0.269 0.2310.967 0.0000.346 F10.812 0.008 0.337 0.2350.652 0.0540.446F10.585 0.000 0.304 0.3640.667 0.0000.409 AUC0.526 0.108 0.506 0.4560.717 0.4690.393AUC 0.437 0.016 0.418 0.451 0.3040.5000.650 BERT0.7800.6660.861BERT0.8120.6260.850 Car Acc0.758 0.497 0.332 0.677 0.642 0.4620.742 Rub. Band Acc0.567 0.5830.7670.817 0.733 0.5000.633 P0.760 0.801 0.733 0.6590.761 0.0000.748P0.567 0.6000.9441.000 0.719 0.0000.611 R0.991 0.438 0.171 0.829 0.762 0.0000.989R0.567 0.500 0.567 0.6330.767 0.0000.733 F1 0.860 0.566 0.278 0.734 0.761 0.0000.852F10.567 0.545 0.7080.7750.742 0.0000.667 AUC0.599 0.527 0.4920.617 0.497 0.3830.575AUC 0.561 0.4630.767 0.2790.783 0.2760.721 BERT 0.8090.6030.737BERT0.8030.6390.784 Caster Wheel Acc0.400 0.833 0.3000.785 0.367 0.4620.583 Screw Acc0.622 0.600 0.444 0.6260.689 0.4600.600 P0.7370.9731.000 0.957 0.818 0.0000.717P0.7600.929 0.619 0.596 0.682 0.4320.700 R0.3110.800 0.067 0.629 0.200 0.0000.733R0.633 0.433 0.433 0.6831.000 0.0590.700 F10.4380.878 0.1250.759 0.321 0.0000.725F10.691 0.591 0.510 0.6610.811 0.0960.700 AUC 0.4680.830 0.533 0.543 0.4100.6090.301AUC0.638 0.620 0.450 0.441 0.298 0.4000.704 BERT0.7540.6260.796BERT0.7790.6570.784 Clip Acc0.6420.622 0.461 0.585 0.344 0.1540.403 Servo Acc0.671 0.333 0.362 0.5080.750 0.4620.447 P0.817 0.703 0.619 0.5651.000 0.0000.634P0.7730.833 0.737 0.527 0.752 0.0000.726 R0.5960.750 0.5001.000 0.017 0.0000.246R0.794 0.139 0.2330.8290.994 0.0000.418 F10.6890.726 0.5530.722 0.033 0.0000.354F10.784 0.238 0.354 0.6440.856 0.0000.530 AUC0.6280.537 0.442 0.518 0.275 0.1800.423AUC 0.479 0.4520.492 0.330 0.342 0.3410.495 BERT0.7980.6400.876BERT0.7970.6380.851 Clock Acc0.568 0.532 0.6440.723 0.356 0.4620.458 Slide Acc 0.647 0.320 0.3870.523 0.200 0.4620.493 P0.669 0.734 0.717 0.6980.857 0.0001.000P0.8021.0000.833 0.542 0.000 0.0000.766 R0.696 0.4660.7700.857 0.041 0.0000.155R0.742 0.150 0.2920.743 0.000 0.0000.536 F10.682 0.5700.7430.769 0.077 0.0000.269F10.771 0.261 0.432 0.626 0.000 0.0000.631 AUC 0.513 0.529 0.5810.766 0.483 0.5000.669AUC0.5340.533 0.529 0.528 0.345 0.2250.415 BERT0.7920.5970.900BERT0.8030.6390.812 Fan Acc 0.642 0.570 0.427 0.569 0.383 0.4620.771 Sph. Bear. Acc0.500 0.867 0.483 0.4670.633 0.5000.517 P0.748 0.8280.846 0.6670.831 0.0000.792P0.5000.893 0.471 0.4170.595 0.0000.509 R0.787 0.537 0.287 0.400 0.220 0.0000.940R0.900 0.833 0.267 0.167 0.833 0.0000.900 F10.767 0.652 0.429 0.500 0.348 0.0000.860F10.6430.862 0.340 0.2380.694 0.0000.651 AUC 0.4090.6020.566 0.472 0.523 0.3890.532AUC 0.5470.894 0.483 0.440 0.629 0.1140.641 BERT0.7970.6300.890BERT0.8260.6620.726 Gear Acc0.740 0.198 0.638 0.739 0.513 0.4620.819 Sticky Roller Acc0.511 0.689 0.5110.7200.756 0.5290.689 P0.801 0.0000.8801.000 0.808 0.0000.797P0.6330.944 0.700 0.6740.788 0.4880.691 R0.897 0.000 0.633 0.514 0.514 0.0000.953R0.633 0.567 0.467 0.7360.867 0.0640.967 F10.847 0.000 0.737 0.679 0.628 0.0000.868F10.633 0.708 0.560 0.7260.825 0.1050.806 AUC 0.418 0.3580.644 0.397 0.451 0.5120.879AUC 0.5170.778 0.533 0.523 0.713 0.4740.811 BERT0.8150.6430.855BERT0.8140.6230.689 Hinge Acc0.6830.700 0.417 0.339 0.550 0.4620.617 Toothpaste Acc0.244 0.589 0.5780.908 0.567 0.4310.656 P0.7601.0000.917 0.409 0.821 0.0000.750P0.2330.700 0.6840.914 0.583 0.4170.667 R0.844 0.600 0.244 0.514 0.511 0.0000.733R0.222 0.311 0.2890.914 0.467 0.1430.622 F10.8000.750 0.386 0.456 0.630 0.0000.742F10.227 0.431 0.4060.914 0.519 0.2130.644 AUC 0.4970.7160.589 0.516 0.544 0.0410.526AUC 0.257 0.535 0.5780.832 0.528 0.5790.689 BERT0.8070.6430.790BERT0.7880.6190.870 Liquid Acc0.5780.7560.822 0.750 0.644 0.6830.667 U Disk Acc0.5040.613 0.442 0.462 0.500 0.4460.512 P0.641 0.828 0.9230.941 0.9381.0000.667P0.5020.599 0.445 0.000 0.000 0.4440.533 R0.833 0.800 0.800 0.533 0.500 0.3671.000R0.9750.683 0.475 0.000 0.000 0.1140.200 F10.725 0.8140.857 0.681 0.652 0.5370.800F10.6630.638 0.460 0.000 0.000 0.1820.291 AUC 0.453 0.7330.833 0.512 0.6910.8270.776AUC 0.5130.585 0.442 0.437 0.404 0.3110.606 BERT0.8170.6400.666BERT0.7850.6500.906 Lock Acc0.678 0.328 0.3670.785 0.333 0.4620.361 Zipper Acc0.4720.706 0.406 0.554 0.350 0.4620.600 P0.682 0.000 0.5420.957 0.000 0.0000.593P0.5900.914 0.7100.875 0.800 0.0000.653 R0.967 0.000 0.3250.629 0.000 0.0000.133R0.683 0.617 0.183 0.200 0.033 0.0000.835 F10.800 0.000 0.4060.759 0.000 0.0000.218F10.6330.736 0.291 0.326 0.064 0.0000.733 AUC 0.442 0.323 0.388 0.231 0.0290.5570.780AUC 0.2620.7910.517 0.477 0.273 0.2240.388 BERT0.7740.6280.909BERT0.8000.6490.711 Avg. Acc0.635 0.591 0.4540.639 0.329 0.4700.592 P0.7130.739 0.717 0.6900.853 0.5000.724 R0.8030.645 0.368 0.580 0.053 0.0500.625 F10.7550.689 0.486 0.630 0.100 0.0910.621 AUC 0.4950.555 0.512 0.503 0.426 0.4560.584 BERT0.7980.6320.803 O-VAD41 Table 9: Results on IPAD [26] across 16 industrial scenarios (S: synthetic, R: real). “†”=trained; “∗”=training-free. For each (scenario, level, metric), the best/2nd- best across methods is highlighted. S01S02S03S04 MethodLv. Acc P R F1 AUC Acc PR F1 AUC Acc PR F1 AUC Acc PR F1 AUC MNAD.p † vid 0.524 0.500 0.200 0.286 0.509 0.368 0.300 0.375 0.333 0.375 0.579 1.000 0.200 0.3330.600 0.368 0.000 0.000 0.000 0.389 frm 0.6140.857 0.029 0.055 0.431 0.7800.386 0.176 0.2420.548 0.739 0.000 0.000 0.000 0.418 0.645 0.000 0.000 0.000 0.419 S3R † vid 0.6190.5620.9000.6920.5090.7370.6670.7500.7060.716 0.421 0.000 0.000 0.000 0.267 0.421 0.000 0.000 0.000 0.133 frm 0.481 0.4230.8820.572 0.467 0.7440.3900.5090.4420.700 0.605 0.261 0.293 0.276 0.443 0.3640.3560.9760.521 0.325 Qwen3 ∗ vid0.7861.000 0.250 0.4000.625 0.500 0.333 0.750 0.462 0.575 0.357 0.143 0.250 0.182 0.3250.643 0.4290.7500.5450.675 frm0.7551.000 0.011 0.0220.5050.837 0.150 0.005 0.010 0.500 0.874 0.000 0.000 0.000 0.4970.814 0.000 0.000 0.000 0.497 GPT-5 ∗ vid 0.643 0.000 0.000 0.000 0.275 0.500 0.286 0.500 0.364 0.4380.643 0.333 0.250 0.2860.6000.714 0.000 0.000 0.0000.625 frm 0.738 0.000 0.000 0.000 0.490 0.693 0.000 0.000 0.000 0.4120.8840.538 0.294 0.3810.6350.819 0.000 0.000 0.0000.500 URF ∗ vid 0.524 0.500 0.600 0.545 0.327 0.6321.000 0.125 0.222 0.443 0.5790.6670.4000.500 0.511 0.5261.000 0.100 0.182 0.311 frm 0.580 0.4760.6810.560 0.469 0.405 0.2440.9460.388 0.528 0.5830.3750.9340.5350.738 0.5400.4170.7490.5360.585 VERA † vid0.714 0.000 0.000 0.000 0.5000.714 0.000 0.000 0.000 0.4500.714 0.000 0.000 0.000 0.525 0.357 0.000 0.000 0.000 0.400 frm0.752 0.000 0.000 0.0000.5000.841 0.000 0.000 0.000 0.5000.879 0.000 0.000 0.000 0.500 0.774 0.000 0.000 0.000 0.473 O-VAD (Ours) ∗ vid 0.524 0.5000.8000.615 0.396 0.474 0.4441.0000.6150.688 0.526 0.5261.0000.690 0.533 0.4210.4710.8000.593 0.389 frm 0.488 0.255 0.158 0.195 0.424 0.391 0.145 0.421 0.216 0.444 0.560 0.3120.5910.409 0.598 0.344 0.254 0.438 0.322 0.370 S05S06S07S08 MethodLv. Acc PR F1 AUC Acc P R F1 AUC Acc P R F1 AUC Acc P R F1 AUC MNAD.p † vid 0.333 0.000 0.000 0.000 0.5000.5260.5001.0000.667 0.378 0.5260.667 0.200 0.308 0.556 0.6320.615 0.800 0.696 0.567 frm 0.7230.914 0.024 0.047 0.389 0.804 0.000 0.000 0.000 0.319 0.831 0.000 0.000 0.000 0.400 0.7940.333 0.005 0.0090.594 S3R † vid 0.4001.000 0.100 0.182 0.240 0.526 0.5001.000 0.667 0.1780.5790.583 0.700 0.636 0.3440.7370.6671.0000.800 0.444 frm 0.5430.3400.6560.4480.583 0.805 0.000 0.000 0.000 0.396 0.171 0.1680.9970.287 0.300 0.351 0.2350.9530.377 0.447 Qwen3 ∗ vid 0.571 0.3330.500 0.4000.550 0.435 0.083 0.333 0.133 0.392 0.357 0.3081.000 0.471 0.550 0.571 0.4001.000 0.5710.700 frm0.898 0.111 0.002 0.004 0.5000.943 0.021 0.001 0.0020.4990.965 0.000 0.000 0.0000.496 0.856 0.000 0.000 0.000 0.496 GPT-5 ∗ vid 0.643 0.400 0.5000.444 0.312 0.429 0.462 0.857 0.600 0.1730.643 0.429 0.750 0.5450.6750.714 0.500 0.750 0.6000.688 frm 0.868 0.110 0.053 0.072 0.503 0.699 0.109 0.081 0.093 0.4570.909 0.000 0.000 0.000 0.4680.8590.487 0.4020.4410.662 URF ∗ vid 0.4671.000 0.200 0.333 0.4000.7370.667 0.8890.7620.667 0.579 0.571 0.8000.667 0.333 0.579 0.556 1.0000.714 0.233 frm 0.549 0.3390.6310.441 0.523 0.4840.2300.7020.3460.493 0.2820.1870.9840.314 0.484 0.285 0.2231.000 0.365 0.424 VERA † vid 0.643 0.000 0.000 0.000 0.338 0.526 0.183 0.126 0.127 0.287 0.429 0.300 0.750 0.429 0.512 0.714 0.000 0.000 0.000 0.500 frm0.896 0.000 0.000 0.000 0.4950.824 0.000 0.000 0.000 0.456 0.882 0.000 0.000 0.000 0.4540.862 0.000 0.000 0.000 0.500 O-VAD (Ours) ∗ vid0.733 0.7141.0000.8330.570 0.421 0.444 0.889 0.5930.794 0.526 0.5261.0000.6900.622 0.526 0.526 1.000 0.690 0.639 frm 0.575 0.325 0.470 0.3850.591 0.5340.1620.3330.218 0.411 0.6060.184 0.394 0.2510.568 0.598 0.208 0.340 0.259 0.500 S09S10S11S12 MethodLv. Acc PR F1 AUC Acc P R F1 AUC Acc P R F1 AUC Acc P R F1 AUC MNAD.p † vid 0.368 0.167 0.125 0.143 0.352 0.421 0.000 0.000 0.000 0.444 0.5261.000 0.100 0.182 0.5500.5790.5561.0000.714 0.367 frm 0.783 0.000 0.000 0.000 0.490 0.763 0.000 0.000 0.000 0.360 0.735 0.000 0.000 0.000 0.452 0.755 0.250 0.001 0.003 0.440 S3R † vid0.6840.6000.7500.667 0.523 0.6320.636 0.700 0.667 0.567 0.421 0.000 0.000 0.000 0.000 0.421 0.000 0.000 0.000 0.144 frm 0.703 0.275 0.248 0.260 0.459 0.4450.2640.7540.3910.519 0.267 0.2650.9990.419 0.329 0.757 0.000 0.000 0.000 0.260 Qwen3 ∗ vid 0.357 0.273 0.750 0.400 0.475 0.571 0.333 0.500 0.400 0.550 0.357 0.3081.000 0.471 0.550 0.286 0.2861.000 0.444 0.500 frm 0.840 0.030 0.002 0.004 0.4960.838 0.000 0.000 0.000 0.4980.928 0.000 0.000 0.000 0.4980.859 0.293 0.008 0.016 0.502 GPT-5 ∗ vid 0.571 0.250 0.250 0.250 0.5000.714 0.500 0.250 0.3330.613 0.286 0.125 0.250 0.167 0.550 0.357 0.308 1.000 0.4710.762 frm 0.793 0.209 0.134 0.163 0.528 0.812 0.006 0.001 0.002 0.482 0.829 0.000 0.000 0.000 0.4450.8720.545 0.4950.5190.727 URF ∗ vid 0.632 0.667 0.250 0.364 0.477 0.5790.5561.0000.714 0.2670.6840.750 0.6000.6670.656 0.474 0.000 0.000 0.000 0.200 frm0.8421.0000.2500.3990.567 0.7640.500 0.020 0.038 0.314 0.7110.454 0.4620.4580.670 0.408 0.2790.903 0.426 0.554 VERA † vid0.714 0.000 0.000 0.0000.6250.714 0.000 0.000 0.0000.6250.714 0.500 0.500 0.5000.6500.643 0.000 0.000 0.000 0.450 frm0.849 0.000 0.000 0.000 0.5000.842 0.000 0.000 0.0000.5000.896 0.000 0.000 0.000 0.481 0.852 0.000 0.000 0.000 0.495 O-VAD (Ours) ∗ vid 0.474 0.4441.0000.6150.580 0.526 0.5261.0000.690 0.433 0.526 0.5261.0000.690 0.467 0.5260.526 1.0000.6900.539 frm 0.5320.2910.8480.4330.654 0.460 0.2410.6000.344 0.438 0.4890.2790.586 0.3780.498 0.5610.3140.6830.4310.611 R01R02R03R04Avg. MethodLv. Acc PR F1 AUC Acc PR F1 AUC Acc PR F1 AUC Acc P R F1 AUC Acc PR F1 AUC MNAD.p † vid 0.600 0.5711.0000.727 0.4290.867 0.8671.0000.929 0.1150.824 0.8750.9330.903 0.300 0.8951.000 0.895 0.9440.500 0.5550.643 0.529 0.581 0.522 frm 0.649 0.000 0.000 0.000 0.535 0.683 0.000 0.000 0.000 0.452 0.5780.500 0.001 0.002 0.483 0.4340.600 0.002 0.004 0.514 0.7160.410 0.006 0.012 0.430 S3R † vid 0.6000.750 0.375 0.500 0.5000.800 0.9170.8460.880 0.500 0.1761.000 0.067 0.125 0.0671.0001.0001.0001.0000.5450.836 0.271 0.4090.539 frm 0.576 0.432 0.776 0.5550.602 0.681 0.403 0.037 0.068 0.366 0.431 0.4261.0000.597 0.460 0.5470.6490.4200.5100.531 0.661 0.3090.166 0.216 0.495 Qwen3 ∗ vid 0.643 0.400 0.500 0.4440.600 0.643 0.333 0.250 0.286 0.525 0.513 0.357 0.333 0.345 0.479 0.357 0.222 0.500 0.308 0.400 0.487 0.268 0.550 0.361 0.507 frm 0.780 0.000 0.000 0.000 0.4970.8620.480 0.010 0.0200.5040.812 0.131 0.003 0.005 0.499 0.791 0.000 0.000 0.000 0.4950.876 0.072 0.002 0.004 0.499 GPT-5 ∗ vid0.714 0.000 0.000 0.000 0.500 0.7861.000 0.250 0.4000.625 0.6431.000 0.286 0.444 0.490 0.714 0.000 0.000 0.0000.6000.607 0.375 0.386 0.380 0.518 frm0.786 0.000 0.000 0.000 0.500 0.832 0.063 0.015 0.025 0.489 0.6631.000 0.128 0.2270.5640.799 0.000 0.000 0.000 0.500 0.815 0.273 0.119 0.1660.531 URF ∗ vid 0.6670.714 0.625 0.667 0.589 0.2001.000 0.077 0.143 0.308 0.706 0.917 0.733 0.815 0.367 0.000 0.000 0.000 0.000 0.190 0.582 0.5900.9290.721 0.417 frm 0.5930.4450.8040.573 0.546 0.6630.4380.2600.326 0.460 0.573 0.4930.3880.434 0.487 0.581 0.5800.9240.712 0.515 0.6710.331 0.1640.220 0.512 VERA † vid0.714 0.000 0.000 0.000 0.500 0.714 0.000 0.000 0.000 0.525 0.695 0.476 0.060 0.1380.490 0.714 0.000 0.000 0.000 0.5000.658 0.238 0.089 0.130 0.519 frm0.786 0.000 0.000 0.000 0.5000.862 0.000 0.000 0.000 0.5000.742 0.000 0.000 0.000 0.5050.799 0.000 0.000 0.000 0.5000.852 0.000 0.000 0.000 0.490 O-VAD (Ours) ∗ vid 0.600 0.5711.0000.7270.679 0.800 0.917 0.846 0.8800.7120.882 0.8821.0000.9380.5670.947 1.0000.9470.973 0.500 0.582 0.5880.9540.7140.565 frm 0.5870.4470.9000.5970.595 0.544 0.3310.4480.3810.512 0.523 0.385 0.215 0.2760.545 0.488 0.587 0.297 0.3950.572 0.515 0.2910.4770.3380.518 42M. Yuan et al. Table 10: Intermediate outputs of O-VAD across the three benchmarks. For each dataset we report source video properties (duration, FPS) and per-stage intermediate outputs: Stage-1 grounded object count and Stage-2 detected state count, each summarized as min. / avg. / max. over the evaluation set. Video InfoIntermediate Results (min. / avg. / max.) DatasetDuration (s)FPS Stage 1: Grounding → #Objects Stage 2: Tracking → #States Phys-AD 0.98 / 2.10 / 4.00601 / 2.31 / 121 / 10.61 / 73 LiquidAD 8.20 / 13.77 / 204.50 301 / 5.44 / 1613 / 104.56 / 416 IPAD1.04 / 22.40 / 40.00 251 / 3.27 / 101 / 29.88 / 134 Table 11: Human evaluation questionnaire. All dimensions use a 1–5 Likert scale (1 = strongly disagree, 5 = strongly agree) except Q10 (preference ranking). Part ID Dimension & DescriptionScale Det. Q1 Binary correctness. Correctly identifies anomalous vs. nor- mal? 1–5 Q2 Type correctness. Predicted type matches the ground-truth defect? † 1–5 Q3 Object identification. Affected object(s) correctly identified? ‡ 1–5 Q4 Temporal localization. Anomaly accurately localized in time? ‡ 1–5 Expl. Q5 Faithfulness. Explanation reflects actual video content, free of hallucinations. 1–5 Q6 Completeness. All salient state changes covered without sig- nificant omissions. 1–5 Q7 Causal coherence. Causal reasoning is logically sound and evidence-consistent. 1–5 Act. Q8 Diagnostic usefulness. Operator could act on the report with- out re-watching the video. 1–5 Ovr. Q9 Overall quality. Overall quality of the anomaly report.1–5 Pref. Q10 Preference. Rank all method reports for the same video for real deployment. Rank 1–3 † Scored only when both ground truth and prediction are anomalous. ‡ Scored only when provided. O-VAD43 Qwen3-VL-32B Answer: “No anomaly detected.” ”prediction": "cot_reasoning": "1. **Lab Equipment Identification**: -The machine appears to be a multi-channel automated pipetting system, likely from ThermoScientific (as indicated by the branding). -It has multiple pipette tips arranged in a row, positioned above a transparent microplate with 96 wells. -There are also blue racks holding pipette tips and possibly reagent reservoirs. -The system moves vertically and horizontally to perform liquid transfers. 2. **Liquid Flow Observation**: -In frame 0.3, the pipette tips are raised above the microplate; no liquid is visible yet. -In frame 1.5, the tips are lowered into a blue liquid reservoir, suggesting aspiration. -In frame 2.7, the tips are moved over the microplate and appear to dispense liquid into the wells —the wells show increasing levels of blue liquid. -In frame 3.8, the tips have been lifted, and the microplate is being moved out of position, indicating completion of dispensing. -No visible spills, bubbles, or irregular liquid distribution are observed across the frames. 3. **Procedural Errors**: -The movement of the pipettor is smooth and controlled. -The tips engage correctly with the reservoir and microplate. -No signs of misalignment, unsteady motion, or incorrect tool usage. 4. **Physical Checks**: -No visible leaks around the pipette tips or reservoirs. -Tips appear intact and properly seated throughout the process. -After dispensing, no residual liquid remains visibly clinging to the tips (in frame 3.8). 5. **Benign Variations Ignored**: -Minor reflections on plastic surfaces and slight camera shake are present but do not affect functionality assessment. 6. **Anomaly Determination**: -All actions appear consistent with standard automated liquid handling procedures. -No anomalies such as cross-contamination, incorrect volume delivery, or mechanical failure are evident. -The entire sequence (frames 0.3–3.8) shows normal operation. : No anomalies detected.", "json_report": "is_anomalous": "No", "reason": "The automated pipetting process proceeds normally with accurate aspiration and dispensing into the microplate. No spills, bubbles, cross-contamination, or mechanical issues are observed..” O-VAD (Ours) Answer: material_anomaly, unintended_single_channel_drip, [30-40]. "report": "anomaly_detected": true, "num_anomalies": 1, "overall_severity": "medium", "overall_confidence": 0.74,"anomalies": ["anomaly_type": "material_anomaly", "anomaly_subtype": "unintended_single_channel_drip", "severity": "low", "description": "At frames 30-40, the pipette head (obj_id: 1) shows material_release where only one tip (noted as the second from the right) releases a droplet while other tips remain unchanged. This pattern suggests unintended dripping/residual droplet rather than a controlled multi-channel dispense step.","affected_objects": ["1"],"evidence_frames": [30,40],"start_frame": 30,"end_frame": 40,"confidence": 0.35960000000000003, "reasoning_trace": "step1_observation": ” Object_1 (obj_id: 1; multi-channel pipette tips/head) shows repeated movement and deformation events across the run. Key events: (a) deformation 0-10 where tips go from bent/wavy to straight (deformation resolved); (b) vertical motion cycles: up 10-20, down 20-30, down/engage 40-60, down 110-120, down 180-190, up 280-290, down 310-320, and stop/engage 340-360 (movement_change). (c) multiple liquid handling events: droplet release from a single tip 30-40; aspiration indicated by blue liquid appearing in tips 120-130; dispensing indicated 130-170 and again 180-210 (material_release). (d) critically, severe deformation is detected at 270-280 (tips go from straight to several visibly bent/curved; severity: severe), followed by severe deformation resolution 290-300 (bent to straight). Additional moderate deformations occur 190-210, ......, 360-380. (single highlighted) (obj_id: 2; plastic; initially intact/attached/unused) shows movement of its rack 0-20, then deformation 20-30 (compressed/irregular top edge), and later multiple deformations 320-370 (slight to moderate bending/compression) consistent with contact/pressure. It is also associated with multiple material_releaseevents into wells (30-60, 110-190, 260-270, 300-310). ...... rack (empty) (obj_id: 4; plastic) tracking includes deformation of a flexible tube/cable 0-10 and 310-350 (bending/tautness changes), plus many movement/material_releaseevents that appear to actually correspond to the pipette head and plates (likely tracking/label noise).” "step2_expectation": "In a normal automated pipetting operation, the pipette head (plastic tips) should move smoothly at steady speed with consistentdown/up cycles, align with wells/reservoirs, and avoid excessive lateral forces. Tips may ...... "step3_comparison": "Two deviations stand out as potentially real process faults: 1) Severe tip bending on the pipette head (obj_id: 1) at frames 270-280 (change_type: deformation, severity: severe): several tips transition from straight/aligned to visibly bent/curved. This exceeds normal minor elastic deflection expected when entering wells or a rack and suggests a collision/misalignment event during engagement. 2) Unintended single-channel droplet release from the pipette head (obj_id: 1) at frames 30-40 (material_release): only one tip (\"second from the right\") releases a droplet while others remain unchanged. In a controlled dispense, either all intended channels dispense or dispensing occurs at a defined step; isolated dripping can indicate a seal/pressure imbalance or residual droplet formation. (e.g., obj_id: 1 at 190-210, ........", "step4_causation": "Plausible causes for the severe pipette tip bending (obj_id: 1, 270-280) include: XY misregistration between pipette head and target (plate/rack), incorrect Z-height calibration causing tips to bottom out, an obstructed well/warped plate causing uneven contact,or a mechanical compliance issue (loose head, worn guide) leading to lateral loading during insertion. droplet at 30-40 could be caused by: residual liquid on one tip from a prior aspiration, partial clog/air leak in one channel causing delayed drip, inconsistent tip fit on that channel, or pressure controltransient. lid flex (obj_id: 3, 320-330) is real, it could be due to the lid catching on a fixture or being contacted by the moving head/gantry during transport..", "step5_classification": "Identified anomalies: -manipulation_failure/ misalignment_collision: severe bending of pipette tips (obj_id: 1) during engagement (270-280). -material_anomaly/ unintended_drip_single_channel: isolated droplet release from one pipette tip (obj_id: 1) at 30-40. lid flex/removal (obj_id: 3, 320-340) is not classified as an anomaly due to insufficient evidence of damage or process failure beyond normal handling/out-of-view removal.”, "step6_severity": "Overall severity is medium. The severe bending event (obj_id: 1, 270-280) indicates a likely collision/misalignment ..." represent real product damage and process/equipment faults, yielding a high-severity anomaly." GPT-5 Answer: “No anomaly detected.” "report": "anomaly_detected": false, "overall_confidence": 0.62, "reason": "reasoning": "step1_observation": ”......(a) a green pipe structure and a metal bracket remaining static throughout, and (b) a white cylindrical object (paint-roller-like) that begins slight movement/rotation, rotates more noticeably, then slows and becomes static again., "step2_expectation": ""For an industrial manipulation operation with a roller-like cylinder mounted in a frame/bracket, normal behavior can include controlled rotation (start/stop) without visible damage, deformation, detachment.......” "step3_comparison": "No deviations can be established against expectations...., the behavior (roller rotates then stops; frame/bracket remain static) is not inherently abnormal and does not indicate damage, leakage, deformation, or a grip/fixture failure.", "step4_causation": "No anomaly-specific causation is supported. The observed start/stop rotation could be due to normal actuation, frictional settling, or a controlled test cycle; none of these imply a flaw or damage ......." "step5_classification": "No anomalies detected; therefore no anomaly classification applies.” "step6_severity": “none. There is no evidence of damage, malfunction, or process failure, and the motion is consistent with normal controlled rotation.”, "summary": ”.....No grounded anomaly can be asserted. The caption describes a white cylindrical roller rotating and then stopping while the green frame and metal bracket remain static, which is consistent with normal controlled manipulation and shows no explicit damage, deformation, detachment, or leakage." O-VAD (Ours) Answer: Yes, The detected anomaly is crushing_of_cardboard_core_and_roll. "report": "anomaly_detected": true, "num_anomalies": 3, "overall_severity": "high", "overall_confidence": 0.82, "anomalies": [ "anomaly_type": "deformation_anomaly","anomaly_subtype": "crushing_of_cardboard_core_and_roll", "severity": "high","description": "Object 1 (cardboard tube) exhibits repeated moderate crushing/flattening (frames 0-10 left-side compression; 60-70 right-side indentation; 140-150 right-side indentation; 160-170 right-end cap indented).Object 3 (paper towel roll) shows repeated moderate compression/bending (110-120, 120-130, 130-140, 150-160).", "affected_objects": [”1”, “3”], "evidence_frames": [0, 10, 60, 70, 110, 120, 130, 140, 150,160, 170], "confidence": 0.85, "reasoning_trace": "step1_observation": "Object 1 (Cardboard tube; material: cardboard; initial: intact) shows repeated deformation events: (a) frames 0-10 deformation with left side compressed/flattened (moderate), (b) frames 60-70 deformation with right side indented/flattened (moderate), (c) frames 120-130 slight compression at top contact area (slight), (d) frames 140-150 right-side indentation (moderate), and (e) frames 160-170 right-end cap indented inward (moderate).Additionally, several deformation events attributed to obj_id=1 describe bending of a 'lower horizontal segment' and 'lower right segment/finger' around frames 90-120, ..., (c) surface_changeframes 160-170: new opening/tear on right end (moderate), and (d) surface_changeframes 170-179: the previously visible tear/opening is no longer visible (moderate), suggesting either partial closure, occlusion, or tracking/visibility change. aligns with the roll (white cylinder) rotating/moving and then becoming static again, while the green structure and metal bracket remain static; the key tracked abnormalities are not motion-related but damage/deformation-related during/after manipulation.", "step2_expectation": "In a controlled industrial manipulation task, a paper towel roll/cardboard tube should be gripped with force ..... Normal behavior would include smooth rotation/translation ..... The robotic gripper (metal structure) should not itself deform under expected loads..... Surface integrity of the paper roll ends should remain intact (no new tears/openings) during routine handling.", "step3_comparison": "Observed behavior deviates from expectations in three main ways: 1) Product damage 2014core/roll crushing: Object 1 shows moderate crushing/flattening early (frames 0-10) and again at multiple later intervals (60-70, 140-150, 160-170), inconsistent with gentle handling of a cardboard tube. ...... 2) Surface tearing: Object 3 develops a new tear/opening at the right end at frames 60-70 and again at 160-170, which is not expected....... 3) Possible end-effector structural/compliance issue: Object 2 is reported as deforming (100-150). A metal gripper should not visibly deform under normal operating forces; this suggests either excessive force, ......", "step4_causation": "Plausible causes include: (a) excessive grip force or incorrect force control parameters causing crushing of the cardboard core (obj 1) and compression/tearing of the paper roll (obj 3); (b) misalignment between gripper and rollleading to point loading at the right end, producing localized indentation and tears ......; (c) mechanical interference/collision with the static green structure/fixture causing the roll to be pinched while rotating; (d) gripper hardware issuesuch as a bent finger, loose linkage, or failed compliant pad leading .......(e) pre-existing weakness in the roll end or core seam that propagates into a tear under normal force.......", "step5_classification": "1) deformation_anomaly: product crushing/indentation of the cardboard tube/paper roll during manipulation (obj 1 and obj 3 across multiple frame ranges). 2) material_integrity_anomaly: tearing/opening at the roll end surface (obj 3 frames 60-70 and 160-170). 3) manipulation_process_anomaly: excessive gripping force / misalignment leading to over-compression and potential end-effector overload (supported by concurrent obj 2 'deformation' flags frames 100-150).", "step6_severity": "Severity is medium to high. Quality impact is significant because the handled item (obj 3 and/or its core obj 1) is visibly deformed and torn, which would likely fail downstream packaging/fit requirements. Task completion may still occur (rotation happens per caption), but the product is damaged......" Fig. 5: Comparative reasoning traces on LiquidAD and Phys-AD. Top: An unintended single-channel drip during automated pipetting. Qwen3-VL-32B observes the full sequence but concludes normal operation, while O-VAD tracks per-tip state changes and identifies the isolated droplet release at frames 30–40. Bottom: A sticky roller undergoes crushing during robotic manipulation. GPT-5 describes the rotation as normal and finds no deviation, while O-VAD tracks progressive deformation of the cardboard core and paper roll across multiple frame intervals, detecting crushing, sur- face tearing, and possible end-effector overload. 44M. Yuan et al. Qwen3-VL-32B Answer: “No anomaly detected.” ”prediction": "cot_reasoning": "1. Objects present: The main object is a servo motor housed in a metallic casing. It has a white plastic cross-shaped output arm attached to its shaft. Inside the casing, blue components (likely circuitry or housing) are visible. A multi-colored wire (red, orange, black) extends from the bottom, indicating power/signal connection. The background is a flat dark surface, likely a workbench. 2. Material estimation: The outer casing appears to be aluminum or metal alloy; the output arm is translucent white plastic; internal parts include blue plastic or coated electronics. Expected behavior: The servo should rotate its output arm smoothly and precisely when commanded, typically oscillating or moving to specific angular positions. 3. Motion dynamics observation: - Frame 0.3s–1.5s: Output arm is stationary at vertical position (one arm pointing down). -Frame 2.7s: Arm rotates rapidly clockwise (~90 degrees), reaching horizontal position (arm pointing right). - Frame 3.8s: Arm returns to original vertical position, appearing stable again. is abrupt but controlled —typical for servo movement under command. No signs of jerking, stalling, or overshoot. 4. Physical checks: -No visible cracks, deformations, or breaks in the plastic arm or casing. -No leakage, discoloration, or smoke. -No excessive vibration or wobbling during rotation. - The arm returns to exact starting position —indicates proper mechanical alignment and no binding. -The servo completes full motion without hesitation —suggests no jamming or electrical issues. 5. Benign variations ignored: Background remains consistent; slight lighting variation due to reflection is normal. 6. Anomaly determination: All observed behaviors align with expected servo operation. The motion is smooth, precise, and repeatable. No physical damage or malfunction detected. : No anomaly present.", "json_report": "is_anomalous": "No", "anomaly_type": "None", "reason": "The servo motor operates normally with smooth and precise rotational motion of the output arm. No physical defects, jams, vibrations, or functional failures are observed.” O-VAD (Ours) Answer: “No anomaly detected.” "report": "anomaly_detected": false, "num_anomalies": 0, "overall_severity": ”none", "overall_confidence": 0.78, "reasoning_trace": [ "step1_observation": "Object 1 (Servo motor horn/servo arm, plastic; initial state: intact, attached) shows a state change labeled as deformation from frames 50-60, but the description indicates it is a pure orientation change/rotation with no shape damage (no bending, stretching, surface damage, or material release). The video caption also describes the servo arm moving clockwise to a diagonal position and later returning to its initial vertical position, consistent with commanded servo motion. 2 (Servo horn attached to servo motor; described as plastic horn with servo body; state changes actually describe the connected wire/cable) shows multiple moderate 'deformation' events that are curvature changes of a flexible wire/cable near the servo exit: frames 10-20 (bent upward), ..........., but the tracked state changes indicate the wire shape changes over time (bending/relaxing). No events indicate tearing, kinking,insulation damage, detachment, or snagging; only bending/curvature changes are reported.......”, "confidence": 1.0 "step2_expectation": "In a controlled industrial manipulation/actuation operation, a plastic servo horn (Object 1) is expected to rotate smoothly when the servo is active and return to a target position without cracking, stripping, or loosening. Orientation change of the horn is normal and should not be treated as physical deformation......”, "confidence": 1.0 "step3_comparison": "Object 1: The frames 50-60 'deformation' is actually rotation/orientation change with explicitly no shape change. This matches expected servo horn motion and is not an anomaly. 2 (wire/cable): Across frames 10-20, 50-60, 70-80, 90-100, and 100-110, the wire curvature changes direction (upward, left, right, then straighter). This is potentially consistent with normal cable compliance during servo actuation. The tracking data does not report sharp kinks, contact damage, or detachment, so there is insufficient evidence that the bending is beyond normal strain. Therefore, no clear deviation from expected behavior is established from the provided events alone.", "confidence": 1.0 "step4_causation": "The observed wire curvature changes (Object 2 events) could be caused by normal mechanical compliance: the servo 2019s internal movement/vibration, slight repositioning of the cable due to gravity, or minor torsion .......", "confidence": 1.0 "step5_classification": "No anomaly is confidently supported by the provided state-change evidence. Object 1’s event is normal rotation. Object 2’s wire bending appears consistent with flexible-cable behavior during actuation and lacks indicators of damage (kink, abrasion, pull-out) .........", "confidence": 1.0 "step6_severity": "Severity: none. The tracked changes correspond to expected servo horn rotation (Object 1, frames 50-60) and plausible flexible wire curvature changes (Object 2, frames 10-20 through 100-110) without evidence of damage, detachment, leakage, or manipulation failure. No impact on task completion or safety is indicated by the events."confidence": 1.0 "state_changes": ["obj_id": "1", "obj_name": "Servo motor horn (servo arm)", "start_frame": 50, "end_frame": 60, "change_type": "deformation", "description": "The highlighted object (a plastic cross-shaped component) has rotated significantly between the two frames. Its shape remains unchanged (no compression, bending, or stretching), but its orientation relative to the background has clearly shifted. There is no evidence of materialrelease, surface damage, or texture change.", "severity": "none (for deformation of shape), but moderate (for orientation change)", "confidence": 0.8......"obj_id": "2", "obj_name": "Servo horn (attached to a servo motor)", "start_frame": 100, "end_frame": 110, "change_type": "deformation", "description": "The object highlighted with the pink contour (a wire or tube) shows a noticeable change in its curvature. In the earlier frame,it is more curved, while in the later frame, it appears straighter and less bent.", "severity": "moderate", "confidence": 0.8],] "summary": "The only change on Object 1 (plastic servo arm) at frames 50-60 is a rotation/orientation shift explicitly without shape damage,which is normal for an actuated servo horn. Object 1's reported 'deformations' are curvature changes of a flexible wire/cable near the servo exitacross frames 10-20, 50-60, 70-80, 90-100, and 100-110; these are physically plausible for a compliant cable during servo operation and show no evidence of kinking, abrasion, detachment, or process failure. No actionable anomaly is supported by the provided tracking data." Qwen3-VL-32B Answer: “Functional Failure -Gripping Mechanism.” ”prediction": “cot_reasoning”: “1. Objects Present and Process: -A red rectangular object (likely a magnet or plastic block) is held by a metallic clamp mechanism attached to a black box labeled “Backyard”. -The clamp appears to be part of a robotic arm or automated system. -In frame 0.3s, .......... -In frame 3.8s, the red object has fallen out of view, and the clamp remains open. 2. Material Estimation and Expected Behavior: -Red object: Likely plastic or rubber-coated magnet; should be rigid but may have slight flexibility. -Clamp: Metallic (aluminum or steel), designed for precise gripping and release. -Expected behavior: The clamp should grip securely and release smoothly without dropping prematurely or failing to hold. 3. State Changes and Motion Dynamics: -From 0.3s to 1.5s: Clamp starts opening —normal. -From 1.5s to 2.7s: Clamp fully opens —normal. -From 2.7s to 3.8s: Red object falls rapidly out of frame —this suggests it was not held securely during release. 4. Physical Checks: -No visible cracks, deformations, or breaks in any components. -No signs of leakage or vibration. -However, the red object drops immediately upon release, which indicates either: -Insufficient friction/grip before release (possibly due to misalignment or wear), -Or the object was never properly secured initially, -Or the release mechanism triggered too early or too abruptly. 5. Benign Variations: ............... ......—specifically, insufficient holding force or improper timing of release. -Since the object is supposed to be held until commanded to release, its immediate drop implies a potential mechanical or control anomaly. : There is an anomaly —likely a functional failure in the gripper’s ability to maintain secure hold prior to intentional release. " "json_report": "is_anomalous": ”Yes", "anomaly_type": " Functional Failure -Gripping Mechanism", "reason": " The red object falls immediately after the clamp begins to open, indicating insufficient grip strength or premature release, suggesting a malfunction in the clamping mechanism.." O-VAD (Ours) Answer: “The detected anomaly is over_clamping_crush_and_surface_marring.” "report": "anomaly_detected": true, "num_anomalies": 3, "overall_severity": ”medium", "overall_confidence": 0.8, "anomalies": [ "anomaly_type": "deformation_anomaly", "anomaly_subtype": "over_clamping_crush_and_surface_marring", "severity": "medium","description": "Obj_id1 (Plastic bottle cap) undergoes repeated compressive deformation while gripped (50-60, 60-70, 70-80, 80-90, 90-100, 110-120) and shows explicit surface damage/indentations at 100-110 (deformation + surface_changewith notches/roughened end). This pattern is inconsistent with controlled non-damaging manipulation of a plastic cap and suggests excessive grip force or poor jaw contact geometry.", "affected_objects": ["1"],"evidence_frames": [50,60,70,80,90,100,110,120],"start_frame": 50, "end_frame": 120,"confidence": 0.82,......."anomaly_type": "manipulation_failure", "anomaly_subtype": "grip_instability_slip_and_drop", "confidence": 0.6084, ......], "reasoning_trace": [ "step1_observation": " Two tracked plastic objects are present: obj_id1 (\"Plastic bottle cap\", initial state: intact/closed, material: plastic) and obj_id2 (\" tube cap\", initial state: intact/closed, material: plastic). For obj_id1, repeated deformation events are detected while in the gripper...... Overall, the tracked state changes show repeated compressive deformation of the plastic part(s) during gripping and a release/drop event around frames 140-150.......”, "confidence": 1.0 "step2_expectation": " In a controlled industrial manipulation operation, a plastic cap (obj_id1) or similar plastic part (obj_id2) should be gripped with sufficient force to prevent slip, but not so much force that it causes visible compression, bending, or surface indentations.Normal behavior would be: stable pose while held, minimal/no permanent deformation for rigid/semi-rigid plastic caps, no surface marring from jaws, and a controlled release/placement ...........”, "confidence": 1.0 "step3_comparison": "Observed behavior deviates from expectations in two main ways. (1) Excessive part deformation/damage: obj_id1 shows repeated moderate deformation from frames 50-120 (50-60, 60-70, 70-80, 80-90, 90-100, 100-110 with surface_change, 110-120), including bending and visible indentations/notches at 100-110, this is inconsistent with a controlled, non-damaging grip on a plastic cap.......", "confidence": 1.0 "step4_causation": "Plausible causes include: (a) Gripper force set too high or lack of force/torque limiting, causing crushing and jaw marks (consistent with repeated deformation and the surface indentations on obj_id1 at 100-110). (b) Incorrect jaw geometry/contact (sharp edges, misaligned prongs) concentrating stress and creating notches/indentations (obj_id1 at 100-110). (c) Mis-pick or off-center grasp leading to bending and eventual slip/rotation (obj_id1 at 70-100 bending and 130-140 tilt). (d) Control/timing issue during transport/release causing premature opening or insufficient holding force at the end, resulting in the drop (obj_id1 at 140-150). (e) Part variability/material defect (softer-than-expected plastic) could amplify deformation, but the repeated pattern across both obj_id1 and obj_id2 points more strongly to process/gripper settings than to a single bad part.”,"confidence": 1.0 "step5_classification": " 1) deformation_anomaly: over-clamping / crushing damage to plastic part(s) (obj_id1 frames 50-120 with surface damage at 100-110; obj_id2 frames 40-90). 2) manipulation_failure: grip instability / slip-then-drop (obj_id1 frames 130-150).", "confidence": 1.0 "step6_severity": " Overall severity is medium. The deformation and surface indentations on obj_id1 (notably 100-110) indicate quality-impacting damage that may render the cap unusable or out-of-spec. The drop event (140-150) can cause additional damage and disrupt downstream handling, but there is no evidence here of catastrophic equipment failure or safety-critical escalatio.....", "confidence": 1.0, "state_changes": ["obj_id": "1", "obj_name": "Plastic bottle cap", "start_frame": 50, "end_frame": 60, "change_type": "deformation", "description": "The pink-highlighted object (red block) is visibly compressed in the second frame compared to the first. The jaws of the gripper have moved closer together, causing the object to deform and flatten slightly along the direction of the applied force.", "severity": "moderate", "confidence": 0.8, "summary": "Tracking data indicates abnormal handling of plastic parts: obj_id1 (plastic bottle cap) experiences repeated moderate deformation while clamped from frames 50-120, including a deformation+surface_changeevent with indentations/notches at 100-110, and then shows grip instability (130-140) followed by a drop (140-150). Obj_id2 also shows repeated compression/bending (40-90), consistent with excessive grip force or misalignment. These deviations suggest over- clamping and a grip stability failure, with medium overall severity due to likely part quality damage and an uncontrolled release.." Fig. 6: Failure cases on servo and magnet categories. Top: A servo motor with a restricted rotation angle is misclassified as normal by both Qwen3-VL-32B and O- VAD. The observed small-angle motion is visually indistinguishable from a normal commanded rotation. Bottom: A degaussed magnet falls off the board. Qwen3-VL-32B incorrectly attributes the anomaly to a gripping mechanism failure. O-VAD detects over-clamping deformation but fails to identify the root cause as loss of magnetism, since the magnet visually resembles an ordinary red plastic block. O-VAD45 Qwen3-VL-32B Answer: “No anomaly detected.” ”prediction": "cot_reasoning": "1. Objects present: A mechanical keyboard switch with a transparent housing, white internal mechanism, and a green keycap. Above it, a metallic actuator or pressing tool with a circular hole is positioned vertically. 2. Material estimation: - The switch housing is likely made of clear plastic (e.g., polycarbonate). -The internal components appear to be plastic and metal (spring and stem). -The keycap is made of rubber or soft plastic. -The actuator above is metallic (likely aluminum or steel). : The actuator presses down on the keycap, compressing the switch’s spring and triggering the internal mechanism (e.g., stem movement, contact closure). Upon release, the spring should return the keycap to its original position. 3. Motion dynamics and state changes: -Frame 0.3s: Actuator is at rest above the keycap, not touching it. -Frame 1.5s: Actuator has descended and makes contact with the keycap; keycap begins to depress. -Frame 2.7s: Keycap is fully depressed under pressure from the actuator. -Frame 3.8s: Actuator retracts upward, and the keycap returns to its original position —indicating normal spring rebound. 4. Physical checks: -No visible cracks, breaks, or deformations in the switch or keycap. -The keycap moves smoothly up and down without sticking or hesitation. -Spring rebound appears consistent and immediate upon release. -No signs of leakage, vibration, or looseness observed. -All components remain aligned throughout the cycle. 5. Benign variations: Background remains dark and uniform; no camera jitter or lighting changes affect interpretation. 6. Anomaly determination: The switch operates as expected —full depression and clean return. No physical or functional anomalies detected.", "json_report": "is_anomalous": "No", "anomaly_type": "None", "reason": "The mechanical switch undergoes full compression and rebounds normally without any visible defects such as sticking, cracking, deformation, or functional failure.” O-VAD (Ours) Answer: “No anomaly detected.” "report": "anomaly_detected": false, "num_anomalies": 0, "overall_severity": "medium", "overall_confidence": 0.74, "reasoning_trace": [ "step1_observation": "The video caption describes a silver metal tool pressing a green keyboard switch keycap down and then retracting, with the keycap returning to its original position (elastic compression/rebound) while the transparent housing remains static. However, the onlytracked object is obj_id=1, described as a Metallic tube (likely a toothpaste or ointment tube), material metal, initial_stateintact, not squeezed. For obj_id=1, repeated state changes of type cdeformationare detected across multiple frame ranges: 0-10 (moderate), 10-20 (moderate), 30-40 (moderate....... 150-160 (slight), and 170-179 (moderate). The deformation descriptions consistently indicate ......”, "confidence": 1.0 "step2_expectation": "In a controlled industrial manipulation/press operation, the pressing tool (a rigid metal component) should not plastically deform under normal operating forces; it should maintain its geometry across cycles. If the pressed item is a keyboard switch/keycap assembly, normal behavior is elastic travel: the keycap moves down and returns, while the housing remains intact and static. For a true metallic tube, some deformation can occur if intentionally squeezed, but that would be a different process than a controlled keycap press; also, repeated progressive indentation at the same edge suggests unintended contact/overload rather than normal handling of a rigid tool.”, "confidence": 1.0 "step3_comparison": "Observed behavior for obj_id=1 deviates from expectations for a rigid pressing tool: multiple moderate deformation events (0-10, 10- 20, 30-40, 40-50, 110-120, 120-130, 140-150, 170-179) indicate the objects lower edge becomes increasingly indented/flattened and its height reduces. This is inconsistent with a normal press where the tool remains dimensionally stable and only the switch/keycap compresses elastically. The deformation is also described as occurring at the contact area with the green object below, suggesting the pressing interface is being damaged or is not sufficiently rigid.", "confidence": 1.0 "step4_causation": "Plausible causes include: (1) excessive press force or over-travel causing the tool edge (obj_id=1) to yield/bend; (2) misalignment leading to edge loading (contact concentrated on a corner/edge rather than a flat face........; (3) incorrect tooling material/part installed (e.g., thin-walled metal component used .....; (4) repeated cycling after initial damage, worsening the deformation over time (consistent with multiple deformation intervals).", "confidence": 1.0 "step5_classification": "deformation_anomaly: tool/fixture damage (plastic deformation of pressing interface) affecting obj_id=1, evidenced by repeated moderate deformation events at the lower edge during contact with the green object.", "confidence": 1.0 "step6_severity": "medium. The deformation appears persistent and repeated (multiple moderate events through frame 179), which can degrade press accuracy, cause inconsistent actuation force/displacement, ....", "confidence": 1.0 "state_changes": ["obj_id": "1", "obj_name": "Metallic tube (likely a toothpaste or ointment tube)", "start_frame": 0, "end_frame": 10, "change_type": "deformation", "description": "The object's lower edge, highlighted in pink, has become noticeably indented and uneven in the later frame compared to the earlier frame, where it was straight and flat. This indicates that the object has been compressed or deformed, likely due to pressurefrom contact with the green object below it.", "severity": "moderate", "confidence": 0.8,........ "obj_id": "1", "obj_name": "Metallic tube (likely a toothpaste or ointment tube)", "start_frame": 170, "end_frame": 179, "change_type": "deformation", "description": "The object highlighted with the pink contour has become more compressed vertically in the later frame. The bottom edge appears to be pushed upward, indicating that the object is being compressed from above. The overall height of the object is reduced, and the lower contour is more indented compared to the earlier frame.", "severity": "moderate", "confidence": 0.8]] "summary": "Tracking data indicates obj_id=1 (metallic component) undergoes repeated moderate deformation at its lower edge (0-10, 10-20, 30-40, 40-50, 110- 120, 120-130, 140-150, 170-179), described as indentation/flattening and vertical compression at the contact area with the greenobject. In a normal controlled press of a switch/keycap, the product should compress elastically while the metal tool remains dimensionally stable; the observed persistent deformation suggests tool/fixture damage due to excessive force, over-travel, or misalignment." Qwen3-VL-32B Answer: “No anomaly detected.” ”prediction": “cot_reasoning”: “1. Objects present: A pink plastic clothespin-like object lying horizontally on a black surface, and a white rectangular object positioned vertically above it, possibly a sensor or part of a testing apparatus. The clothespin appears to be open, with its two arms spread apart. 2. Material estimation: The pink object is likely made of plastic, given its smooth texture and matte finish. The white object appears to be rigid, possibly plastic or metal, with a circular hole that may serve as a lens or sensor aperture. 3. Motion dynamics: There is no observable motion across the frames. All objects remain static throughout the sequence. No interaction occurs between the white object and the pink clothespin. 4. Physical checks: -Stuck: Not applicable; no moving parts are engaged. -Leakage: Not applicable; no fluids or gases involved. -Deformation: The pink object shows no signs of bending, warping, or structural distortion. -Breakage: No visible cracks or broken components. -Functional failure: Since there’s no movement or activation, we cannot assess functionality directly. However, if this were intended to close or clamp, the lack of motion suggests possible non-functionality —but without context, this is speculative. - Vibration: None observed. -Loose: The pink object lies flat and stable; no indication of looseness. -Crack: No visible fractures or splits. 5. Benign variations: Background remains uniformly black, lighting consistent, no camera jitter or noise detected. 6. Anomaly determination: Given the static nature of all objects and absence of any visible damage or unexpected behavior, there is no evidence of an anomaly. The scene appears to be a controlled setup for observation or testing, not a malfunctioning system. : No anomaly detected." "json_report": "is_anomalous": "No", "anomaly_type": "None", "reason": "All objects remain static with no visible damage, deformation, or functional issues. The scene shows no signs of physical anomalies such as cracks, breaks, leaks, or vibrations." O-VAD (Ours) Answer: “No anomaly detected.” "report": "anomaly_detected": false, "num_anomalies": 0, "overall_severity": ”none", "overall_confidence": 0.62, "reasoning_trace": [ "step1_observation": "Tracked objects: obj_id1 = plastic bottle cap with dispensing hole (initial_state: opened, facing downward, plastic). obj_id2 = plastic spring clamp (initial_state: opened, plastic with visible metal spring). Video caption describes a white smartphone moving downward to press on the top jawof a clothespin/clamp, causing it to open further under load, then the phone lifts and the clamp returns to its prior position. There are no explicit tracked transitions.....”, "confidence": 1.0 "step2_expectation": "In a normal controlled manipulation/pressing operation with a plastic spring clamp (obj_id2) and a rigid pressing tool/object, expected behavior is: (a) clamp jaws elastically open under applied force, (b) metal spring stores energy without slipping out of seat, (c) upon release, clamp returns to its original open/closed resting state ......if it is not involved in contact, it should remain static.”, "confidence": 1.0 "step3_comparison": "No deviations can be confirmed from the tracking evidence because there are no detected state change events and no frame ranges provided to verify irreversible deformation, cracking, spring displacement, or grip/press failure. ......", "confidence": 1.0 "step4_causation": "Since no confirmed deviation is present in the state-change log, there is no anomaly to attribute. If an anomaly had been present in such a scenario, plausible causes would include excessive force causing plastic cracking, misalignment causing torsional loading andhinge damage, or spring seat failure leading to spring pop-out; but none of these are evidenced by the provided tracking/state-change data.”,"confidence": 1.0 "step5_classification": "No anomaly classification is warranted based on the provided tracked-object metadata and the absence of any detected state change events indicating damage or malfunction.", "confidence": 1.0 "step6_severity": "medium. The deformation appears persistent and repeated (multiple moderate events through frame 179), which can degrade press accuracy, cause inconsistent actuation force/displacement, ....", "confidence": 1.0 "state_changes": ["obj_id": "1", "obj_name": "Metallic tube (likely a toothpaste or ointment tube)", "start_frame": 0, "end_frame": 10, "change_type": "deformation", "description": "The object's lower edge, highlighted in pink, has become noticeably indented and uneven in the later frame compared to the earlier frame, where it was straight and flat. This indicates that the object has been compressed or deformed, likely due to pressurefrom contact with the green object below it.", "severity": "moderate", "confidence": 0.8,........ "obj_id": "1", "obj_name": "Metallic tube (likely a toothpaste or ointment tube)", "start_frame": 170, "end_frame": 179, "change_type": "deformation", "description": "The object highlighted with the pink contour has become more compressed vertically in the later frame. The bottom edge appears to be pushed upward, indicating that the object is being compressed from above. The overall height of the object is reduced, and the lower contour is more indented compared to the earlier frame.", "severity": "moderate", "confidence": 0.8]] "summary": "Tracking data indicates obj_id=1 (metallic component) undergoes repeated moderate deformation at its lower edge (0-10, 10-20, 30-40, 40-50, 110- 120, 120-130, 140-150, 170-179), described as indentation/flattening and vertical compression at the contact area with the greenobject. In a normal controlled press of a switch/keycap, the product should compress elastically while the metal tool remains dimensionally stable; the observed persistent deformation suggests tool/fixture damage due to excessive force, over-travel, or misalignment." Fig. 7: Failure cases on button and clip categories. Top: A stuck button that fails to actuate is misclassified as normal by both methods. O-VAD detects repeated deformation of the pressing tool but attributes it to tool/fixture damage rather than recognizing the button’s inability to depress. Bottom: A clip that cannot be pressed shows no observable state changes, leading both methods to predict normal behavior despite the ground-truth anomaly. 46M. Yuan et al. Table 12: Human and LLM-as-judge evaluation results (mean scores, 1–5 scale). Human: N=5 evaluators× 10 videos; LLM: GPT-4o judge× 10 videos. URF-ZS: URF- ZS-HVAA [24]; Qwen3: Qwen3-VL-32B [1]. Q4 is excluded from LLM evaluation due to the absence of temporal ground truth. Best per row per evaluator is bold; ties share bold. Human EvaluationLLM-as-Judge Part DimensionURF-ZS Qwen3 O-VAD URF-ZS Qwen3 O-VAD Det. Q1. Binary correctness1.242.704.981.402.604.60 Q2. Type correctness1.082.604.261.001.80 1.80 Q3. Object identification1.143.944.461.002.603.20 Q4. Temporal localization 1.082.323.96N/A Detection avg.1.142.894.421.132.333.20 Expl. Q5. Faithfulness1.043.284.101.304.903.90 Q6. Completeness1.003.364.581.204.003.90 Q7. Causal coherence1.003.024.301.203.703.60 Explanation avg.1.013.224.331.234.203.80 Act. Q8. Diagnostic usefulness1.002.844.161.102.303.40 Ovr. Q9. Overall quality1.043.004.281.202.703.50 Pref. Q10. Win rate (%)0.018.082.010.030.070.0