HiProbe-VAD: Tuning-Free Video Anomaly Detection

• The paper introduces HiProbe-VAD, a tuning-free framework that leverages hidden-state probing in pre-trained MLLMs for video anomaly detection using minimal few-shot labels.
• It employs Dynamic Layer Saliency Probing (DLSP) to automatically select the most discriminative intermediate layer based on divergence, discriminative ratio, and entropy metrics.
• The framework achieves competitive performance on UCF-Crime and XD-Violence benchmarks by integrating lightweight anomaly scoring with adaptive temporal localization.

Searching arXiv for the specified paper and closely related entries to ground the article in current literature. arxiv_search(query="(Cai et al., 23 Jul 2025)", max_results=5, sort_by="submittedDate")

HiProbe-VAD is a framework for video anomaly detection (VAD) that uses hidden-state probing in pre-trained multimodal LLMs (MLLMs) without requiring fine-tuning. It is introduced in "HiProbe-VAD: Video Anomaly Detection via Hidden States Probing in Tuning-Free Multimodal LLMs" (Cai et al., 23 Jul 2025). The method is motivated by the observation that intermediate hidden states of MLLMs contain information-rich representations and exhibit higher sensitivity and linear separability for anomalies than the output layer. On that basis, HiProbe-VAD selects a single informative intermediate layer through Dynamic Layer Saliency Probing (DLSP), applies a lightweight anomaly scorer to the selected hidden states, and then performs adaptive temporal localization to detect anomalous segments and generate explanations (Cai et al., 23 Jul 2025).

1. Problem setting and methodological position

Video anomaly detection aims to identify and locate deviations from normal patterns in video sequences. In the formulation used by HiProbe-VAD, the practical constraints of the field are twofold: substantial computational demands and a reliance on extensive labeled datasets. HiProbe-VAD addresses these constraints by operating with frozen pre-trained MLLMs and a small labeled subset rather than full model fine-tuning (Cai et al., 23 Jul 2025).

The framework is explicitly positioned as tuning-free rather than fully unsupervised. That distinction is technically important. DLSP is run on a small labeled subset of training videos, described as approximately $1\%$, and the anomaly scorer is trained offline on few labeled samples extracted from the selected layer. A common misunderstanding would be to equate the absence of fine-tuning with the absence of supervision; HiProbe-VAD does not fine-tune the backbone MLLM, but it does use few-shot labels for layer selection and for the lightweight classifier (Cai et al., 23 Jul 2025).

Within the paper’s comparative taxonomy, HiProbe-VAD is evaluated against weakly supervised, unsupervised, tuning-free, and fine-tuned MLLM methods. This positioning matters because the framework is intended to extract anomaly-relevant structure from pre-trained MLLM internals rather than to adapt the backbone through end-to-end optimization.

2. Dynamic Layer Saliency Probing

The core contribution of the method is Dynamic Layer Saliency Probing (DLSP), which identifies the single intermediate layer whose hidden states best separate normal and anomalous content. Let an input video $V$ be tokenized and passed through a pre-trained MLLM with $L$ layers. For each layer $l$, the hidden-state vector is denoted

$\mathbf{h}^{(l)} \in \mathbb{R}^D,$

taken immediately after the visual-text fusion and just before the feed-forward sub-layer (Cai et al., 23 Jul 2025).

DLSP assigns to each layer a saliency score $S(l)$. The score is constructed from three statistics computed over the small labeled subset.

First, the method computes the average per-dimension Kullback–Leibler divergence between the normal-sample distribution $\mathcal{N}(\mu^N_{l,d},(\sigma^N_{l,d})^2)$ and anomalous-sample distribution $\mathcal{N}(\mu^A_{l,d},(\sigma^A_{l,d})^2)$: $$D_{KL}(l)=\frac1D\sum_{d=1}^D \tfrac12\Bigg[ \ln\frac{(\sigma^A_{l,d})^2}{(\sigma^N_{l,d})^2} +\frac{(\sigma^N_{l,d})^2+(\mu^N_{l,d}-\mu^A_{l,d})^2}{(\sigma^A_{l,d})^2} -1\Bigg].$$

Second, it computes the Local Discriminant Ratio: $$\mathrm{LDR}(l)=\frac1D\sum_{d=1}^D \frac{(\mu^N_{l,d}-\mu^A_{l,d})^2} {(\sigma^N_{l,d})^2+(\sigma^A_{l,d})^2+\epsilon}.$$

Third, it computes the Shannon entropy under uniform binning: $V$0 where $V$1 is the empirical probability of $V$2 falling in bin $V$3.

Each metric is then z-score normalized across layers: $V$4

The final saliency score is

$V$5

and the selected layer is

$V$6

All downstream computations use $V$7. The design implies that anomaly sensitivity is treated as a property of internal representational geometry rather than only of the model output. This suggests a probing-based view of MLLMs in which anomaly cues are more salient at an intermediate processing stage than at the terminal prediction layer.

3. Lightweight scoring and temporal localization

After selecting $V$8, HiProbe-VAD performs inference by uniformly sampling key frames from each video. The implementation uses $V$9 frames per 24-frame clip. Each sampled frame $L$0 is processed by the frozen MLLM to obtain

$L$1

A logistic-regression-style anomaly scorer is then applied: $L$2 The parameters $L$3 are trained offline on the few labeled samples from layer $L$4 by minimizing binary cross-entropy,

$L$5

with $L$6, and the implementation uses an L-BFGS solver for this small-scale problem (Cai et al., 23 Jul 2025).

Temporal localization operates on the per-frame anomaly score sequence $L$7. The raw scores are first smoothed with a Gaussian kernel to reduce jitter. Adaptive thresholding is then applied using

$L$8

where $L$9 are few-shot statistics and $l$0 is a threshold factor. Consecutive frames above $l$1 are grouped as anomalous events, while frames below $l$2 form normal intervals. In the implementation, the Gaussian smoothing width is $l$3 and the threshold factor is $l$4 (Cai et al., 23 Jul 2025).

The paper also states that, once anomalous segments are localized, the MLLM auto-regression can produce concise textual descriptions such as “person starts fighting” and “crowd disperses quickly” for the detected segments.

4. Datasets, implementation, and evaluation protocol

HiProbe-VAD is evaluated on two standard benchmarks: UCF-Crime and XD-Violence (Cai et al., 23 Jul 2025).

Dataset	Composition and split	Evaluation
UCF-Crime	1,900 untrimmed videos (128 h), 13 anomaly classes, split 1,610 train / 290 test	Frame-level AUC (ROC)
XD-Violence	4,754 videos (217 h), 6 violence types, split 3,954 train / 800 test	Average Precision (AP)

The backbone MLLMs used in the experiments are InternVL2.5, Qwen2.5-VL, LLaVA-OneVision, and Holmes-VAU. DLSP is run on an approximately $l$5 few-shot subset, and the logistic regression is trained on hidden-state dimension $l$6. The hardware specification given is an NVIDIA RTX 4090 (Cai et al., 23 Jul 2025).

These implementation details situate the method in a regime of frozen large-scale multimodal backbones combined with a very small learned head. A plausible implication is that the computational burden is shifted away from end-to-end adaptation and toward representational selection and lightweight post hoc scoring.

5. Quantitative performance and ablation evidence

The reported benchmark results compare HiProbe-VAD against weakly supervised, unsupervised, tuning-free, and fine-tuned MLLM baselines (Cai et al., 23 Jul 2025).

Dataset	Method / Backbone	Score
UCF-Crime	S3R / I3D	85.99 AUC
UCF-Crime	CLIP-TSA / ViT	87.58 AUC
UCF-Crime	DYANNET / I3D	84.50 AUC
UCF-Crime	LAVAD / ViT	80.28 AUC
UCF-Crime	VERA / ViT	86.55 AUC
UCF-Crime	HiProbe-VAD / InternVL2.5	86.72 AUC
UCF-Crime	HiProbe-VAD / Qwen2.5-VL	85.89 AUC
UCF-Crime	HiProbe-VAD / LLaVA-OV	82.26 AUC
UCF-Crime	Holmes-VAU / ViT	88.91 AUC
XD-Violence	UR-DMU / I3D	81.66 AP
XD-Violence	CLIP-TSA / ViT	82.19 AP
XD-Violence	LAVAD / ViT	62.01 AP
XD-Violence	HiProbe-VAD / InternVL2.5	82.15 AP
XD-Violence	HiProbe-VAD / Qwen2.5-VL	80.94 AP
XD-Violence	HiProbe-VAD / LLaVA-OV	76.32 AP
XD-Violence	Holmes-VAU / ViT	89.51 AP

The paper characterizes these results as showing that HiProbe-VAD outperforms existing training-free and most traditional approaches. On UCF-Crime, the best tuning-free result reported for HiProbe-VAD is 86.72 AUC with InternVL2.5; on XD-Violence, the best tuning-free result is 82.15 AP with InternVL2.5 (Cai et al., 23 Jul 2025).

The ablation study on the InternVL2.5 backbone is central to the method’s internal validation. Removing DLSP by fixing layers drops performance by up to 8.1 percentage points, with “Layer 16” fixed yielding AUC 78.6. Replacing logistic regression with SVM or distance scoring loses approximately 2 percentage points and approximately 6 percentage points, respectively. Fixed thresholding at 0.25 or 0.75 can cost 10–16 percentage points, which the paper interprets as evidence for the need for adaptive temporal localization (Cai et al., 23 Jul 2025).

6. Cross-model transfer, zero-shot behavior, and interpretive significance

A prominent claim of the framework is cross-model generalization without tuning. In the zero-shot protocol reported, the system is trained on UCF-Crime and tested on XD-Violence, and vice versa. The results are as follows (Cai et al., 23 Jul 2025).

Zero-shot backbone	UCF-Crime AUC (%)	XD-Violence AP (%)
LLaVA-OneVision	76.21	71.17
Qwen2.5-VL	80.60	75.86
InternVL2.5	81.35	77.04
Holmes-VAU	86.33	85.65

These figures support the paper’s claim that the selected intermediate hidden states encode transferable anomaly cues even without fine-tuning. The qualitative analysis further states that HiProbe-VAD-generated temporal curves align closely with ground-truth abnormal events, and that the MLLM can generate concise textual descriptions for each detected segment (Cai et al., 23 Jul 2025).

From a broader methodological perspective, HiProbe-VAD treats anomaly detection as a probing problem over latent multimodal representations. The framework can therefore be summarized in three steps stated explicitly in the paper: it dynamically identifies and probes the single best intermediate layer of a frozen MLLM, trains a tiny anomaly scorer on those hidden states with only a few labels, and enforces temporal consistency by adaptive smoothing and thresholding. In that formulation, its contribution lies not in modifying the backbone, but in extracting and exploiting a layer whose hidden states are most discriminative for abnormality.