Training-free Video Anomaly Detection

• Training-free Video Anomaly Detection is an approach that identifies deviations in video scenes using statistical models and pretrained multimodal networks without any parameter learning.
• It employs techniques such as order-independent discriminative frameworks, adaptive online learning, and prompt-based vision-language scoring to robustly detect anomalies.
• These methods achieve competitive benchmark performance while offering enhanced interpretability and real-time application potential in diverse surveillance and edge scenarios.

Training-free Video Anomaly Detection refers to the class of methodologies for identifying temporal regions or frames in video streams where the observed patterns diverge from a model of “normal” activity, without any parameter learning or fine-tuning on the current or similar domains. These approaches circumvent the costs and fragilities associated with supervised or unsupervised training on large video datasets, leveraging instead statistical modeling, pretrained foundation models, architectural priors, or even online adaptation. The past five years have seen a rapid proliferation of training-free VAD pipelines, especially those built upon large vision–LLMs (VLMs) and multimodal LLMs (MLLMs) (Zanella et al., 2024, Li et al., 26 Oct 2025, Shao et al., 17 Apr 2025, Cai et al., 19 Dec 2025, Cai et al., 23 Jul 2025, Cai et al., 27 Feb 2026, Bekit et al., 3 Apr 2026). This survey reviews the algorithmic archetypes, model architectures, evaluation standards, and experimental findings in modern training-free VAD research, encompassing both statistical and MLLM-based (zero-shot) paradigms.

1. Problem Setting and Motivation

Traditional video anomaly detection frameworks typically presume access to either: - Extensive labeled data (supervised/weakly-supervised), - Curated background (normal-only, one-class), - Domain-specific unsupervised training (reconstruction, density estimation, discriminative learning).

However, in practical scenarios such as open-world surveillance, crowd monitoring, or edge deployment, labeled “normal” or “abnormal” samples are costly or impossible to obtain, and real-world domain drift invalidates precise feature distributions. Training-free VAD seeks to: - Eliminate all offline parameter learning or supervision for each deployment setting. - Exploit pretrained foundation models’ generic visual–textual world knowledge or statistical structure. - Provide robust, generalizable, and often interpretable anomaly detection across heterogeneous datasets and anomaly types (Zanella et al., 2024, Cai et al., 19 Dec 2025, Bekit et al., 3 Apr 2026, Li et al., 26 Oct 2025, Cai et al., 23 Jul 2025).

2. Classical Statistical and Discriminative Frameworks

Early approaches to training-free anomaly detection in videos were rooted in statistical density modeling or discriminative learning directly on the test video. Notable paradigms include:

• Order-independent discriminative frameworks: Anomalies are those frames or segments that can be linearly separated from others within the same video. The method of (Giorno et al., 2016) leverages repeated random permutations of frame order, splitting each permutation into “past” and “window” sets, and fitting L2-regularized logistic regression to discriminate between them. The average confidence of each frame over many such splits yields a robust anomaly score, independent of the original temporal ordering.
• Adaptive online incremental learning: The “Look at Adjacent Frames” method (Ouyang et al., 2022) maintains a pixel-wise MLP (SIREN) whose parameters are continuously updated online to minimize prediction error on the current frame conditioned on temporal-frequency features. Anomalies are detected as abrupt increases in the per-frame mean squared error under this always-adapting model.
• k-Nearest Neighbor with memory cleansing: In the CKNN pipeline (Yi et al., 2024), instead of learning a discriminative model, precomputed features (CLIP embeddings or motion descriptors) are stored in a “memory bank.” Anomaly scores for each object in a test frame are functions of their k-nearest-neighbor distances to this bank. A key innovation is the one-pass “cleansing” of the memory bank to remove high-density anomalous clusters that would otherwise pollute unsupervised training with spurious “normal” regions.

These methods often serve as baseline comparators, with frame-level AUC performance rates for statistical or discriminative pipelines now surpassed by recent VLM- and MLLM-based techniques (Zanella et al., 2024, Li et al., 26 Oct 2025, Shao et al., 17 Apr 2025, Cai et al., 23 Jul 2025, Cai et al., 19 Dec 2025).

3. VLM/MLLM-Based Architectures: From Vision–Language Prompting to Intermediate State Probing

Modern training-free VAD architectures increasingly exploit the zero-shot semantic alignment, compositional reasoning, and cross-modal world knowledge encoded in foundation models trained on massive image-text corpora. This broad class decomposes into several technical strategies:

• Text Caption + LLM Scoring: LAVAD (Zanella et al., 2024) operates by generating frame-level captions using a frozen vision-LLM (e.g., BLIP-2), applying cross-modal retrieval to select the most semantically consistent captions, and then using a LLM to assign anomaly likelihoods based on summary prompts and context templates. Video-level or windowed summaries are constructed, scored, and then refined by aggregating over semantically similar segments via a cross-modal retrieval step.
• Hierarchical Segmentation and Event Reasoning: VADTree (Li et al., 26 Oct 2025) introduces a hierarchical granularity-aware tree, partitioning video via a Generic Event Boundary Detection (GEBD) model and fusing priors (scene, object, action) with VLMs for event-wise captioning. Each event node is scored using LLM-based anomaly prompts and weighted fusion, enabling multi-scale anomaly localization and explainability while drastically reducing redundant sampling.
• Graph and Temporal Dynamics with MLLMs: EventVAD (Shao et al., 17 Apr 2025) first builds a dynamic spatiotemporal graph over video frames with appearance and motion features, applies graph attention, statistically detects event boundaries, and finally applies a hierarchical prompting strategy to a frozen MLLM for temporally-consistent segment-level anomaly scoring.
• Hidden State Probing and Head Selection in MLLMs: HiProbe-VAD (Cai et al., 23 Jul 2025), HeadHunt-VAD (Cai et al., 19 Dec 2025), and SteerVAD (Cai et al., 27 Feb 2026) demonstrate that intermediate hidden states (layers or attention heads) within MLLMs encode greater anomaly sensitivity and linear separability than final outputs. HiProbe-VAD’s DLSP mechanism computes per-layer saliency metrics to select and probe the most anomaly-discriminative layer, then employs a lightweight logistic regression scorer and adaptive thresholding for temporal grouping. HeadHunt-VAD identifies a robust, sparse subset of experts heads using a multi-criteria analysis (KL divergence, LDA, MMD, NMI), maximizing anomaly separability and stability across prompts. SteerVAD further introduces a meta-controller to actively steer the latent representation manifolds via dynamic, context-dependent scaling, amplifying anomaly signals and suppressing inherited pretraining biases.
• Prompt-adaptive and Query-centric Reasoning: QVAD (Bekit et al., 3 Apr 2026) frames VAD as a dynamic, question-centric dialogue between a lightweight VLM and an LLM agent. The agent iteratively refines queries based on visual context and prior responses, enabling small models (down to 2 B parameters) to rival the detection performance of much larger systems while maintaining real-time operational speed on resource-constrained hardware.
• Object-event Symbolic Reasoning: VED-SR (Zeng et al., 9 Feb 2025) combines off-the-shelf open-vocabulary object detectors with LLM-guided symbolic logic search. Features extracted from detected entities are fed to a symbolic regression engine (augmented with LLM guidance), producing interpretable logical rules for anomaly detection that generalize across domains and detectors.

4. Inference Pipelines and Common Architectural Components

While components vary by method, common structural motifs include:

Pipeline Stage	Role	Representative Approaches
Video segmentation	Partition into clips, windows, or event nodes	LAVAD, VADTree, HiProbe-VAD, EventVAD
Feature extraction	Visual/motion/semantic representations	CLIP/BLIP-2, Mask R-CNN, RAFT, DWT
Pretrained model use	Frozen backbone for encoding/generation	BLIP-2, Qwen2.5-VL, InternVL, LLaVA
Proposal cleaning	Noise reduction, cross-modal or pseudo-anomaly filtering	CKNN, LAVAD
Intermediate probing	Hidden layer/head selection and use	HiProbe-VAD, HeadHunt-VAD, SteerVAD
Temporal reasoning	Smoothing, adaptive/local thresholding, graph propagation	HiProbe-VAD, EventVAD, LAVAD, VADTree
Anomaly scoring	Lightweight classifier or LLM prompt-based estimation	HeadHunt-VAD, LAVAD, QVAD
Explanation	In-context, natural-language event descriptions	HiProbe-VAD, VADTree, QVAD, LAVAD

Temporal localization modules, such as adaptive Gaussian smoothing plus mean–std thresholding (Cai et al., 23 Jul 2025), hierarchical event-tree boundary segmentation (Li et al., 26 Oct 2025), or statistical signal ratio/MAD-based methods (Shao et al., 17 Apr 2025), are viewed as essential for practical VAD.

5. Evaluation Protocols and Experimental Findings

Evaluation of training-free VAD benchmarks centers on large-scale, real-world datasets (UCF-Crime, XD-Violence, UBNormal, ComplexVAD), following standardized metrics:

• Frame-level AUC (area under ROC curve)
• Average Precision (AP)
• AUROC for event or segment-level tasks

Notable quantitative results (reported as state-of-the-art for fully tuning-free methods):

Method	UCF-Crime (AUC %)	XD-Violence (AP %)	UBNormal (AUC %)	Reference
HeadHunt-VAD	87.03	82.6	–	(Cai et al., 19 Dec 2025)
HiProbe-VAD	86.72	82.15	–	(Cai et al., 23 Jul 2025)
SteerVAD	87.15	83.02	–	(Cai et al., 27 Feb 2026)
QVAD	84.28	68.53	79.6	(Bekit et al., 3 Apr 2026)
VADTree	84.7	67.8	65.8	(Li et al., 26 Oct 2025)
EventVAD	82.03	64.04	–	(Shao et al., 17 Apr 2025)
LAVAD	80.28	62.01	64.23	(Zanella et al., 2024)
ZS-CLIP baseline	53.16	17.8	–	(Zanella et al., 2024)

Ablation studies across these works consistently demonstrate the importance of intermediate representation probing, multi-criteria hidden-head selection, and temporal adaptation mechanisms. For example, fixed-thresholding in HiProbe-VAD drops AUC on UCF-Crime from 86.7 % to 70.4 %; removing explainability or adaptive head selection similarly reduces performance (Cai et al., 23 Jul 2025, Cai et al., 19 Dec 2025, Cai et al., 27 Feb 2026).

6. Interpretability, Limitations, and Future Directions

A defining attribute of training-free VAD, especially in the MLLM context, is the capacity for explanation and interpretability: - Many frameworks generate natural-language event descriptions for detected anomalies, allowing human operators to contrast “normal” and “abnormal” intervals (Cai et al., 23 Jul 2025, Li et al., 26 Oct 2025, Bekit et al., 3 Apr 2026, Zanella et al., 2024). - Intermediate state and head probing methods localize attention to spatially precise anomaly regions, confirmed by t-SNE and activation visualizations (Cai et al., 19 Dec 2025, Cai et al., 27 Feb 2026).

Limitations reported by the literature include: - Residual sensitivity to prompt design and semantic recall of the pretrained models (Zanella et al., 2024, Li et al., 26 Oct 2025, Zeng et al., 9 Feb 2025). - Performance degradation in domains where anomalies are subtle, intent-based, or manifest only in rare or ambiguous contexts (Cai et al., 27 Feb 2026, Cai et al., 19 Dec 2025). - Occasional lack of statistical significance testing and potential computational overhead from repeated large-model inference (Zanella et al., 2024, Li et al., 26 Oct 2025, Shao et al., 17 Apr 2025).

Future research directions emphasize: - Plug-and-play adaptation of interpretability modules to new domains with minimal or online calibration (Cai et al., 19 Dec 2025, Cai et al., 23 Jul 2025, Cai et al., 27 Feb 2026). - Deeper integration of object/event reasoning, symbolic logic, and lightweight dialogue-based LLMs for edge deployment (Zeng et al., 9 Feb 2025, Bekit et al., 3 Apr 2026). - Streaming and online/real-time anomaly detection with ultra-low latency constraints (Ouyang et al., 2022, Bekit et al., 3 Apr 2026). - Joint integration of additional modalities (audio, sensor, context) and continual learning mechanisms to improve open-set recognition (Li et al., 26 Oct 2025, Cai et al., 27 Feb 2026, Bekit et al., 3 Apr 2026).

7. Contextual Significance and Outlook

The field of training-free video anomaly detection is notable for driving forward robust, interpretable detection architectures that forgo the need for training on every domain or environment. Evaluation across a standardized suite of difficult, real-world benchmarks confirms that carefully designed pipelines—leveraging pretrained multimodal foundation models, event-structured temporal segmentation, and explainable decision processes—match or exceed the performance of many supervised and unsupervised learning approaches, despite the absence of domain-specific parameter updates (Cai et al., 19 Dec 2025, Cai et al., 23 Jul 2025, Cai et al., 27 Feb 2026, Li et al., 26 Oct 2025, Zanella et al., 2024).

A plausible implication is that, as pretrained foundation models increase in capacity, diversity, and multimodality, training-free anomaly detection pipelines will further close the gap to fully supervised systems, especially in challenging, resource-constrained, or privacy-sensitive deployment contexts. Interpretability and calibration (e.g., through multi-prompt or hierarchical reasoning) remain key differentiators between competitive pipelines. The shift to training-free VAD marks a significant paradigm change, emphasizing efficient, generalizable, and explainable anomaly detection as core desiderata for next-generation video analysis systems.