Efficient Zero-Shot Video Anomaly Detection via Signal-Aware Adaptive Fusion (SAAF)

Efficient Zero-Shot Video Anomaly Detection via Signal-Aware Adaptive Fusion (SAAF) Khaled Waleed Dawoud Muhammad Haris Khan Mohamed bin Zayed University of Artificial Intelligence United Arab Emirates

arXiv:2610.32486v1 [cs.CV] 23 Oct 2026

{khaled.dawoud, muhammad.haris}@mbzuai.ac.ae

Dense

Abstract Deploying Vision-Language Models for video anomaly detection confronts a fundamental constraint. Their semantic demands computational budgets incompatible with the scale of real-world video streams. We present Signal-Aware Adaptive Fusion (SAAF), a framework that achieves efficient zero-shot anomaly detection through three key innovations. First, we introduce a dual-signal adaptive sampler operating in semantic embedding space: a delta signal captures sudden events via frame-to-frame cosine distance, while a drift signal detects static and gradual anomalies through deviation from rolling temporal context. Their ORgated combination with adaptive EMA-based thresholding provides comprehensive anomaly coverage without training data. Second, we propose a threshold-gated cascaded architecture where lightweight frame-level scoring precedes expensive video-level temporal analysis, invoking it only when confidence warrants additional context. Third, we design a signal-aware weighted fusion mechanism that dynamically adapts score combination based on which signal triggered frame selection. Across four benchmarks, SAAF delivers competitive accuracy at substantially lower cost. Among training-free zero-shot methods, SAAF sets a new state of the art on XD-Violence (92.07%), MSAD (86.70%), and Shang (80.54%) AUC-ROC, with XD and MSAD also surpassing supervised baselines. SAAF sustains 44.95, 48.61, and 41.06 FPS on XD, UCF, and MSAD, a 2.43×, 2.63×, and 2.22× throughput improvement over the prior best zero-shot method. Code: https://github. com/kdawoud91/SAAF.

1. Introduction Video anomaly detection (VAD) spans applications from public safety monitoring to industrial hazard oversight [7, © 2027 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses.

Uniform

SAAF selected Anomaly Region

Anomaly

Skipped

(a) Dense

30/30

(b) Uniform

10/30

(c) SAAF

7/30

θ

δ + drift signal 1

6

11

16

Frame Index

21

26

30

Figure 1. Comparison of sampling strategies. (a) Dense sampling. (b) Uniform sampling selects at fixed intervals. (c) SAAF adaptively selects within the same uniform stride pool, concentrating frames around anomalous regions (pink), avoiding redundant frames. SAAF achieves efficient coverage of anomalous content while significantly reducing the number of processed frames.

16, 26]. Yet anomalies are inherently rare and unpredictable. Unlike object detection [5, 23], where categories can be enumerated, what constitutes “abnormal” shifts across contexts and environments. This challenge has driven interest in zero-shot (ZS) approaches that detect anomalies without training on target categories. VisionLanguage Models (VLMs) such as VideoLLaMA [40], Qwen-VL [4], and GPT-4V [1] offer promising capabilities from describing visual scenes, reasoning about behaviors, and assessing normalcy. In principle, a VLM could serve as a universal anomaly detector. However, in practice, computational cost presents a significant obstacle. One hour of 30 FPS video produces 108,000 frames, and even efficient VLMs such as [35] require hundreds of milliseconds per frame, rendering dense analysis impractical. A more useful question is where in a video the VLM’s capacity is actually needed. Anomalies occupy a small fraction of most video footage, and within an anomalous segment, frames vary in how much they contribute to a correct decision. If computation could be steered toward the anomalous frames and the reasoning depth that each video actually requires, the VLM budget would shrink while concentrating on the frames that drive the decision. Existing pipelines offer no such mechanism, since frame selection, frame scoring, and temporal reasoning are each invoked on

fixed intervals [14, 32, 37, 38] with no mechanism to redirect effort based on what the video looks like. Notably, anomalies are not homogeneous in their temporal signature. Sudden events such as explosions, shootings, and road accidents manifest as abrupt departures between adjacent frames. Static or gradually evolving anomalies, such as a slowly developing fire, produce frame-to-frame change yet remain abnormal relative to the surrounding context. The abnormality in the first case lies in what changes while in the second, in what persists. Uniform stride sampling is poorly suited to either regime. By allocating computation at fixed intervals, it processes redundant content while offering no mechanism to concentrate frames where either signature occurs. This motivates our central question: how can adaptive sampling cover both abrupt and gradual anomalies while preserving efficiency on normal content? We propose a dual-signal adaptive sampler combining complementary signals. The delta signal measures frameto-frame change, capturing sudden events. The drift signal measures deviation from a rolling temporal context, detecting divergence from recent patterns. Both operate on semantic embeddings, and support zero-shot operation through vision-language alignment, see Fig. 1. Beyond frame selection, existing VLM-based systems typically run both frame-level and video-level analysis unconditionally for every sampled frame [38], regardless of whether temporal context is actually needed. Our threshold-gated cascaded architecture addresses this inefficiency. Among the anchor frames retained by the dualsignal sampler, we first compute frame-level scores and invoke video-level analysis only on those whose frame-level score exceeds a confidence threshold. This two-level gating, adaptive selection followed by score-based cascading, sharply reduces the number of expensive video-level VLM calls while focusing temporal reasoning where it is most likely to add discriminative value. For frames that trigger both analyses, naive score averaging ignores that different anomaly types benefit differently from each level. Our key insight is that the triggering signal itself reveals the anomaly type. Delta-triggered frames correspond to sudden events where frame-level analysis is most informative. Drift-triggered frames correspond to static or gradual anomalies where video-level context clarifies ambiguity. When both signals fire, equal weighting is applied. This yields our signal-aware weighted fusion mechanism, which dynamically adjusts fusion weights based on which signal caused frame selection. We summarize our key contributions as follows: • Dual-signal adaptive sampling. We introduce complementary temporal signals in semantic embedding space: a delta signal capturing sudden transitions via frame-toframe cosine distance, and a drift signal detecting static and gradual anomalies through deviation from rolling

temporal context. Their OR-gated combination with adaptive EMA thresholding provides coverage of distinct anomaly types without training data. • Threshold-gated cascading. Lightweight frame-level scoring gates expensive video-level analysis, invoking temporal reasoning only when confidence warrants it. • Signal-aware weighted fusion. The triggering signal informs score combination. Delta-triggered frames weight frame-level scores higher, drift-triggered frames weight video-level scores higher. This couples selection rationale to adaptive fusion strategy without additional parameters. • Efficiency and accuracy. Among training-free zeroshot methods, SAAF sets a new SOTA on XD, MSAD, and Shang, with XD and MSAD also surpassing supervised baselines. SAAF sustains amortized throughputs of 44.95, 48.61, 41.06 and 33.55 FPS on XD, UCF, MSAD and Shang delivering 2.43×, 2.63×, 2.22× and 1.81× speedups over the prior best zero-shot method [14].

2. Literature Review Traditional VAD methods, including reconstruction-based autoencoders [7, 11, 18], future-frame prediction [16, 21, 25], and distance-based outlier detection [20, 29], operate as black boxes that produce anomaly scores without explaining what is anomalous or why. These approaches typically require training data and can be sensitive to distribution shifts across scenes and anomaly categories. Vision-Language Models (VLMs) [3, 13, 22, 40] and Large Language Models (LLMs) [1, 10, 28] have enabled zero-shot anomaly detection with semantic interpretability. LAVAD [38] pioneered this direction with a training-free captioner-LLM-scorer pipeline. However, its uniform sampling is content-blind and it unconditionally invokes both frame-level and temporal-level analysis for every sampled frame, processing redundant normal content while failing to concentrate computation on anomalies. Recent work explores multi-stage reasoning for zero-shot detection. Lin et al. [14] propose a unified framework that chains temporal detection with spatial localization and textual understanding, using score-based gates to trigger additional reasoning for ambiguous predictions. However, their approach retains limitations: initial scoring operates on uniformly sampled frames, anomaly context is extracted after scoring rather than guiding frame selection, and reasoning refinement treats all anomaly types identically despite their differing analysis requirements. Table 1 summarises the sampling strategies adopted by recent zero-shot and weakly-supervised VAD pipelines. The dominant choice is a fixed stride, typically every sixteen frames [14, 32, 38], with the remaining methods substituting equally content-independent heuristics, fixed segment lengths [2], fixed-rate sampling within sliding windows [37], or event-uniform sampling after a separate de-

SAAF selected

Uniform

Dense

Anomaly

Skipped

Legend δ Sudden ρ Gradual τ Threshold

(a) Dense

(b) Uniform

S Dense Score

(c) SAAF

(EMA) Exponential

10

5

1

1 Dual-Signal Selection CLIP ViT-B/32

Semantic Embedding

2 Frame-Level Scoring VLM

V

Qwen3-VL-8B

15

3 Conditional Video-Level Fusion

ρt

Delta Signal Frame-to-frame distance

Drift Signal Context deviation

"The image shows a view from the front of a motorcycle, with the handlebars and part of the dashboard visible in the foreground. The road ahead is a two-lane road with a solid white line dividing the lanes.

OR Gate Adaptive Thresholding τ = μ + N · σ (EMA)

A = {anchor frames}

L

LLM Qwen2.5-7B

S frame ∈ [0,1]

4

Propagation & Refinement

Threshold Gate: G(t) = [s frame > τ ] f

<τ

δt

moving average

25

Anomaly region 20

Refined

>τ VideoLLaMA3 + Qwen2.5-7B

Frame-only s=s frame

Signal-Aware Weighted Fusion

λ (S(t)) =

𝜆𝛿

δ triggered

𝜆𝜌

ρ triggered

0.5

both signals

sfused = λ·s frame + (1-λ)·s vid

M-stride Dense Output S = {s1, s 2, ..., sT }

Figure 2. Overview of the SAAF pipeline. Stage 1 extracts delta (frame-to-frame change) and drift (context deviation) signals from CLIP embeddings via adaptive EMA-based thresholding. Stage 2 generates frame-level anomaly scores through VLM captioning and LLM reasoning. Stage 3 conditionally invokes video-level temporal analysis for high-scoring frames, combining frame and video scores via signal-aware weighted fusion. Stage 4 propagates and refines sparse anchor scores to produce dense frame-level predictions.

tector [24]. Across the table, no method allows the video content to drive selection at the first stage which is where the largest computational savings are available, since every frame eliminated here removes one downstream VLM call.

scenarios without per-domain adaptation.

More broadly, routing computation by content is a long-standing principle in efficient video understanding. AdaFrame [34] learns a frame-selection policy, SCSampler [12] forwards only salient clips to a 3D backbone, AR-Net [19] predicts per-frame input resolutions, FrameExit [9] exits early from intermediate features, and LiteEval [33] gates between coarse and fine networks. SAAF is distinguished from prior work by two design choices. First, their saliency signal is learned under closed-set supervision, a policy or gate is trained against action labels that define what counts as informative. Zero-shot VAD has no such labels, so the first-stage signal must be derived intrinsically rather than learned. Second, their notion of saliency is single-mode a unified score of how much a frame contributes to a classification objective which suffices when the target event has a consistent temporal signature but fails when the same pipeline must surface both abrupt and persistent phenomena. We address both by replacing the selector with statistical deviation in a frozen semantic embedding space. The delta signal measures frame-to-frame change in CLIP space, while the drift signal measures deviation from a rolling temporal context. The result is a content-aware stage that requires no training data, properties that make the same selector applicable across traffic, industrial, and crowd

Method

Sampling Strategy

AnyAnomaly [2] EventVAD [24] Holmes-VAU [41] VERA [37] STPrompt [32] LAVAD [38] Unified [14]

Segment length 24, 4 frames per segment Stride 16 per event (MLLM) Stride 16, then 16 frames via ATS 8 frames per 10 s window Stride 16 (UCF), stride 4 (Shang) Stride 16 Stride 16

SAAF (Ours)

Dual-signal δ∨ρ on CLIP, EMA-thresholded

Table 1. Sampling strategies adopted by ZS & WS VAD methods.

3. Methodology We propose SAAF, a zero-shot framework for efficient and robust video anomaly detection. Given a video V = {f1 , f2 , . . . , fT } consisting of T frames, our objective is to produce a dense anomaly score sequence S = {s1 , s2 , . . . , sT }, where each st ∈ [0, 1] represents the likelihood of anomalous activity at frame t. The framework addresses three fundamental challenges: (i) identifying where to allocate VLM computation, (ii) determining when temporal reasoning is necessary, and (iii) deciding how to fuse multi-level predictions. To this end, SAAF consists of four stages: (1) dual-signal adaptive

frame selection, (2) frame-level anomaly scoring, (3) conditional video-level fusion, (4) score propagation and refinement. Figure 2 illustrates the overall pipeline.

3.1. Dual-Signal Adaptive Frame Selection Applying VLM inference densely over all frames is computationally prohibitive and often redundant. Uniform sampling reduces cost but remains content-blind. We therefore aim to identify a sparse set of informative anchor frames that are likely to contain anomalous content. Formally, we seek an indicator function I : {1, . . . , T } → {0, 1} such that: \mathcal {I}(t) = \begin {cases} 1 & \text {if frame } f_t \text { warrants VLM analysis} \\ 0 & \text {otherwise} \end {cases}

(1)

where ⊮[·] is the indicator function and τδ,t is computed from running statistics as described in Section 3.1.4. However, delta alone is insufficient. Anomalies that persist without motion produce low δt despite being semantically abnormal. 3.1.3. Drift Signal: Deviation from Temporal Context To capture persistent and gradually evolving anomalies, we introduce the drift signal. We maintain a rolling temporal anchor over the preceding W frames: a_t = \frac {1}{W} \sum _{i=t-W}^{t-1} e_i

The normalized anchor and drift distance are defined as: \hat {a}_t = \frac {a_t}{\|a_t\|_2}, \quad \rho _t = 1 - \hat {a}_t \cdot e_t

A key motivating our design is that anomalies are not homogeneous: different anomaly types exhibit distinct temporal signatures. Sudden events manifest as abrupt visual transitions, whereas static or gradual anomalies evolve slowly or persist without noticeable frame-to-frame change. Capturing both requires complementary signals operating in semantic embedding space. 3.1.1. Semantic Frame Embedding For each frame ft , we extract a semantic embedding using a pretrained visual encoder ϕ: e_t = \frac {\phi (f_t)}{\|\phi (f_t)\|_2} \in \mathbb {R}^d

(2)

By default, we adopt CLIP ViT-B/32 [22] as our encoder ϕ due to its strong alignment between visual and semantic concepts. The ℓ2 -normalization ensures ∥et ∥2 = 1, enabling cosine similarity as a unified distance metric: \text {sim}(e_i, e_j) = e_i \cdot e_j = \cos (\theta _{ij})

(3)

where θij is the angle between embeddings.

3.1.4. Adaptive Thresholding Fixed thresholds are brittle across diverse scenes with varying baseline dynamics. We therefore employ adaptive thresholds based on running statistics computed via exponential moving average (EMA). For each signal x ∈ {δ, ρ}, we maintain online estimates (t) of the first moment (mean) µx and second moment (mean (t) of squares) νx : \mu _x^{(t)} &= (1 - \alpha )\mu _x^{(t-1)} + \alpha x_t \\ \nu _x^{(t)} &= (1 - \alpha )\nu _x^{(t-1)} + \alpha x_t^2

(10) where α ∈ (0, 1) controls the adaptation rate and xt denotes the value at frame t. The running variance is derived as: (11)

The adaptive threshold follows a sigma-rule formulation:

We define the delta signal as the cosine distance between consecutive embeddings: (4)

This signal captures abrupt semantic transitions, with δt ∈ [0, 2] where δt = 0 indicates identical frames and δt = 2 indicates maximally dissimilar frames. In practice, consecutive video frames yield δt ∈ [0, 1] under normal conditions. A spike in δt indicates rapid change in scene composition or activity. A frame is selected when δt exceeds an adaptive threshold τδ,t : \mathcal {I}_\delta (t) = \mathbb {1}[\delta _t > \tau _{\delta ,t}]

(7)

Unlike delta, which captures instantaneous change, drift measures cumulative deviation from recent temporal context. A frame is selected when drift exceeds its adaptive threshold: \mathcal {I}_\rho (t) = \mathbb {1}[\rho _t > \tau _{\rho ,t}] (8)

(\sigma _x^{(t)})^2 = \nu _x^{(t)} - (\mu _x^{(t)})^2

3.1.2. Delta Signal: Instantaneous Semantic Change

\delta _t = 1 - e_t \cdot e_{t-1} = 1 - \cos (\theta _{t,t-1})

(6)

(5)

\tau _{x,t} = \mu _x^{(t)} + N \cdot \sigma _x^{(t)}

(12)

where N controls detection sensitivity. Higher N raises the threshold and reduces the selection rate. We treat N as a sensitivity parameter (see Supp. A.2 for a sweep across N ∈ {0.5, 1.0, 1.5, 2.0}). Statistic Freeze. When a frame is selected, we suspend updates to the running statistics: (\mu _x^{(t)}, \sigma _x^{(t)}) = \begin {cases} (\mu _x^{(t-1)}, \sigma _x^{(t-1)}) & \text {if } \mathcal {I}(t) = 1 \\ \text {EMA update} & \text {otherwise} \end {cases}

(13)

This mechanism prevents anomalous observations from corrupting baseline statistics.

3.1.5. OR-Gated Selection and Signal Attribution

3.3.1. Threshold-Gated Temporal Reasoning

The final selection indicator combines both signals via logical disjunction:

We apply video-level analysis selectively based on framelevel confidence. Let τf denote the fusion threshold. Videolevel reasoning is triggered when:

\mathcal {I}(t) = \mathcal {I}_\delta (t) \lor \mathcal {I}_\rho (t) = \mathbb {1}[\delta _t > \tau _{\delta ,t}] \lor \mathbb {1}[\rho _t > \tau _{\rho ,t}] (14) This OR-gated design ensures complementary coverage across anomaly types. We additionally define a signal attribution function S : A → {delta, drift, both}:

\mathcal {S}(t) = \begin {cases} \texttt {delta} & \text {if } \mathcal {I}_\delta (t) = 1 \land \mathcal {I}_\rho (t) = 0 \\ \texttt {drift} & \text {if } \mathcal {I}_\delta (t) = 0 \land \mathcal {I}_\rho (t) = 1 \\ \texttt {both} & \text {if } \mathcal {I}_\delta (t) = 1 \land \mathcal {I}_\rho (t) = 1 \end {cases}

(15)

\mathcal {G}(t) = \mathbb {1}[s_t^{\text {frame}} > \tau _f]

(19)

For frames where G(t) = 1, we extract a temporal clip Ct = {ft−k , . . . , ft , . . . , ft+k } and process it through a video-language model Vvid followed by a language model: s_t^{\text {vid}} = \mathcal {L}(\mathcal {V}_{\text {vid}}(C_t, p_{\text {vid}}), p_{\text {score}})

(20)

This achieves reduction in computational cost while focusing temporal reasoning on high-confidence candidates. Critically, δ and ρ are not anomaly categories but orthogonal modes of temporal change. An anomalous event might exhibit an abrupt departure from the immediate past, a sustained deviation from the rolling context, or both. The ORgate covers all three regimes.

3.3.2. Signal-Aware Weighted Fusion For frames receiving both scores, we adapt fusion weights based on the triggering signal S(t). We define the signal-dependent weight function λ mapping {delta, drift, both} → [0, 1]:

3.2. Frame-Level Anomaly Scoring Selected frames undergo semantic analysis via a two-stage VLM-LLM pipeline. For each selected frame ft where I(t) = 1, we compute an anomaly score through visual captioning followed by language-based reasoning. 3.2.1. Visual Captioning

\lambda (\mathcal {S}(t)) = \begin {cases} \lambda _\delta & \text {if } \mathcal {S}(t) = \texttt {delta} \\ \lambda _\rho & \text {if } \mathcal {S}(t) = \texttt {drift} \\ \lambda _{\text {both}} & \text {if } \mathcal {S}(t) = \texttt {both} \end {cases}

where λδ , λρ , λboth ∈ [0, 1] are fusion weights. The fused anomaly score is computed as a combination:

We employ a vision-language model V to generate a natural language description:

s_t^{\text {fused}} = \lambda (\mathcal {S}(t)) \cdot s_t^{\text {frame}} + (1 - \lambda (\mathcal {S}(t))) \cdot s_t^{\text {vid}}

c_t = \mathcal {V}(f_t, p_{\text {cap}})

The scoring function integrating all conditions is:

(16)

where pcap is the captioning prompt. The prompt instructs the model to describe the scene with emphasis on actions, and context (see Supplementary Section C for exact prompts). 3.2.2. Language-Based Anomaly Reasoning

(17)

where pscore is the scoring prompt. The prompt instructs the model to assess whether the scene description depicts normal or anomalous behavior and output a score in [0, 1] (see Supplementary Section C). The complete frame-level scoring function is: s_t^{\text {frame}} = \begin {cases} \mathcal {L}(\mathcal {V}(f_t, p_{\text {cap}}), p_{\text {score}}) & \text {if } \mathcal {I}(t) = 1 \\ 0 & \text {otherwise} \end {cases}

s_t = \begin {cases} s_t^{\text {fused}} & \text {if } \mathcal {I}(t) = 1 \land \mathcal {G}(t) = 1 \\ s_t^{\text {frame}} & \text {if } \mathcal {I}(t) = 1 \land \mathcal {G}(t) = 0 \\ 0 & \text {if } \mathcal {I}(t) = 0 \end {cases}

(22)

(23)

3.4. Score Propagation

The generated caption is evaluated by a large language model L to produce an anomaly score: s_t^{\text {frame}} = \mathcal {L}(c_t, p_{\text {score}}) \in [0, 1]

(21)

(18)

3.3. Conditional Video-Level Fusion While frame-level analysis captures visually salient anomalies, certain events require temporal context for disambiguation. We introduce conditional video-level reasoning that incorporates temporal context when warranted.

Since VLM inference is performed only at selected frames, we propagate scores temporally to obtain dense predictions. Let M denote the maximum propagation stride and next(t) the next anchor after t ∈ A. For each anchor with computed score st , we define the propagation operator: \begin {aligned} \mathcal {P}(s_t) &= \{(t + i, s_t) : i \in \{0, 1, \ldots , \ell _t - 1\}\}, \\ \ell _t &= \min (\mathrm {next}(t) - t,\, M) \end {aligned}

(24)

The dense score sequence is constructed as: \mathcal {S} = \bigcup _{t \in \mathcal {A}} \mathcal {P}(s_t)

(25)

Frames outside any propagation window receive a default fill value sd : s_j = \begin {cases} s_t & \text {if } \exists \, t \in \mathcal {A} : j \in [t,\, t + \ell _t) \\ s_d & \text {otherwise} \end {cases}

(26)

(a) XD (6 anomaly categories)

(b) UCF-Crime (13 anomaly categories)

Figure 3. Per-category AUC-ROC across SAAF pipeline stages. Stage 2 (blue): frame-level scoring. Stage 3 (red): λ-weighted video-level gating. Stage 4 (green): refinement. All categories benefit from multi-stage fusion. Table 2. VAD Performance comparison across XD, UCF, MSAD and Shang. ✓ / ✗ indicate whether a method is zero-shot and training-free in terms of model parameters. Our SAAF results are mean ± std over 3 runs. Method

Zero-shot Training-free

XD [30]

UCF [26]

MSAD [43]

Shang [17]

AUC(%)

AUC(%)

AUC(%)

AUC(%)

Sultani et al. [26] FusedVision [7] RTFM [27] GODS [29] MGFN [6] STPrompt [32] OVVAD [31] Holmes-VAU [41] MULDE [20] UR-DMU [42] AnomalyRuler [36] VERA [37]

✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗

✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✗ ✓ ✓

– – 61.56 – – – – – – – – 88.26

77.92 – – 70.46 86.98 88.08 86.40 88.96 – – 86.55

– – 85.67 – 85.0 – – – – 74.30 – –

– 83.58 – – – – – – 81.40 – 85.20 –

LLAVA-1.5 [15] Eventvad [24] AnyAnomaly [2] MCANet [8] LAVAD [38] Unified-reasoning [14]

✓ ✓ ✓ ✓ ✓ ✓

✓ ✓ ✓ ✓ ✓ ✓

79.62 87.51 – 87.43 85.36 91.34

72.84 82.30 80.70 82.47 80.28 84.28

– – – – – 85.90

– – 79.70 – – –

Ours (adaptive)

✓

✓

92.07±0.02 82.59±0.05 86.70±0.05 80.54±0.03

4. Experiments 4.1. Implementation Details Frame Sampling and Embedding Extraction. We apply stride-16 sampling for signal computation, reducing the candidate pool to approximately 6.25% of original frames. Each sampled frame is encoded via CLIP ViT-B/32, yielding 512-dimensional ℓ2 -normalized embeddings. The dualsignal adaptive sampler operates on these embeddings with EMA-based adaptive thresholding (decay factor α = 0.05), selecting frames whose signal values exceed N = 1 standard deviation above the running mean for VLM inference. A warmup period of 10 frames initializes the running statistics before detection begins. The rolling window for drift computation spans W = 20 frames. To preserve sensitivity during anomaly bursts, EMA statistics are frozen whenever a frame is selected, preventing anomalous observations from corrupting the baseline distribution. Frame-Level Scoring. Selected anchor frames undergo visual captioning via Qwen3-VL-8B-Instruct [35], followed by anomaly reasoning through Qwen2.5-7B-Instruct to produce scores in

[0, 1]. Video-Level Fusion. For frames whose frame-level score exceeds the threshold τf = 0.4, we extract a temporal clip of 10 frames centered at the anchor and process it through VideoLLaMA3-7B [39]. The resulting video-level caption is scored by Qwen2.5-7B-Instruct using the same anomaly prompt. Frame-level and video-level scores are combined via signal-aware weighted fusion as described in Section 3.3, with fusion weights set to λδ = 0.6, λρ = 0.4, and λboth = 0.5. Score Propagation and Refinement. Anchor scores are propagated forward with a maximum stride M = 64; each anchor’s score fills subsequent frames until either the next anchor or M frames, whichever comes first. Frames where neither δ (instantaneous change) nor ρ (contextual deviation) exceeded their respective adaptive thresholds remain unselected and receive a default score of sd = 0.3, encoding a soft normality prior. All three defaults were fixed a priori. N = 1 balances selection rate and AUC-ROC (Supp. A.2), sd = 0.3 is a conservative normality prior, τf = 0.4 is a low threshold chosen to preserve recall. Following [14], We then apply Gaussian smoothing with σ = 100. We then apply Gaussian smoothing on the dense per-frame score sequence with (≈ 3.3 s at 30 fps).

4.2. Results & Discussion Table 2 compares SAAF against supervised and zero-shot methods across four benchmarks. XD. SAAF reaches 92.07% AUC-ROC, surpassing the previous best zeroshot result of 91.34% from Unified [14] while processing 30.31% fewer frames. UCF. SAAF reaches 82.59% AUC-ROC, outperforming LAVAD [38], AnyAnomaly [2], and Eventvad [24] (80.28–82.30%) and remaining competitive with Unified [14] (84.28%) despite processing 23.48% fewer frames. Supervised methods Holmes-VAU [41] and STPrompt [32] score higher (88.96, 88.08%), but benefit from target-domain training and uniform processing. MSAD. SAAF reaches 86.70% AUC-ROC, the highest among supervised (RTFM [27]: 85.67%, MGFN: 85.0%) and zero-shot (Unified: 85.90%) methods, with 34.87% fewer frames processed. Shang. SAAF reaches 80.54% AUC-ROC competitive with supervised MULDE [20]

2)

3)

Deadpool.2.2018__#-51-20

About.Time.2013__#32-31

1) Casino.Royale.2006__#19-20_

Table 3. Per-component computational cost across benchmarks.

# 310

# 785

Frame # 50

c)

1.0 0.8 0.6 0.4 0.2 0.0

# 1140

0 200 400 600 800 1000 1200

0 200 400 600 800 1000 1200

Frame # 90 # 300

1.0 0.8 0.6 0.4 0.2 0.0 1.0 0.8 0.6 0.4 0.2 0.0

# 660

# 1660

1.0 0.8 0.6 0.4 0.2 0.0

0 200 400 600 800 1000 1200

0 200 400 600 800 1000 1200

XD

UCF

MSAD Shang

Stage 1: Signal Extraction (candidate frames) CLIP ViT-B/32 24 0.14 146,449 69,634 14,360 2,550 0

0 200 400 600 800 1000 1200

0 200 400 600 800 1000 1200

b)

# 701

1.0 0.8 0.6 0.4 0.2 0.0

1.0 0.8 0.6 0.4 0.2 0.0

1.0 0.8 0.6 0.4 0.2 0.0

# 660

1.0 0.8 0.6 0.4 0.2 0.0 1.0 0.8 0.6 0.4 0.2 0.0

0

500

500

1000

1000

1500

1500

2000

2000

Stage 2–3: Frame-Level Scoring (selected frames) VLM (Qwen3-VL- 24 8.19 102,066 53,283 9,353

1,898

8B) LLM (Qwen2.5-7B) 24 0.64 102,066 53,283

1,898

9,353

frame

0

500

1000

1500

2000

Figure 4. Shows Qualitative results. Each column shows: (a) signal activation patterns with ground-truth anomaly regions shaded in pink, blue vertical lines indicate delta-only activations, green lines indicate drift-only activations, and orange lines indicate frames where both signals are activated; (b) frame-level anomaly scores; and (c) video-level scores after threshold-gated fusion.

(81.40%) and FusedVision [7] (83.58%). Shang anomalies are brief, limiting the temporal resolution available to δ and ρ under stride-16 sampling. Two further analyses support these results. The percategory breakdown (Fig. 3) confirms the dual-signal taxonomy. Abrupt events (fighting, explosions, shootings) score well at the frame level, while gradual ones (abuse, robbery) gain substantially from video-level fusion. Smoothing applied directly to Stage 2 (Table 7) leaves a consistent margin to the full pipeline across benchmarks, showing Stage 3’s semantic evidence drives the gain.

4.3. Qualitative results Figure 4 Video 1: an explosion spans frames 200–600. Signals (a) concentrate selection within the anomalous region, frame-level scores (b) rise sharply at onset and stay elevated, and gated fusion (c) yields a temporally coherent detection. Video 2: three separate violent events. SAAF samples around each and skips the gaps (a); frame-level scores (b) peak at all three locations and fusion (c) sharpens the contrast against normal segments. Video 3: a normal video with continuous motion drives persistent drift activation (a), yet SAAF selects only 43% of frames. Both scoring stages (b, c) remain low throughout, avoiding false positives under high scene dynamics – detailed success and failure cases are provided in (Supplementary Sec. B).

5. Computation Reduction Analysis Figure 5 shows the per-video distribution of VLM inference calls saved relative to uniform stride-16 sampling across the four benchmarks. Two observations stand out. First, adaptive selection achieves non-trivial reductions across the ma-

Stage 4: Video-Level Fusion (gated anchors, s 15,968 3,300 15,968 3,300

VideoLLaMA3-7B 1 0.84 LLM (Qwen2.5-7B) 24 0.64

14.4 44.95

Total (h) Amortized FPS

6.35 48.61

> τf ) 2,327 487 2,327 487

1.54 41.06

0.33 33.55

jority of videos, with the bulk concentrated in the 20–60% range. XD peaks at 50–60% (138 videos, the modal bin), while UCF and MSAD spread more uniformly across 20– 50%. Second, the long tail extending to 80–90% reduction shows that videos with predominantly static content yield substantial savings without sacrificing detection capability. Table 3 summarizes the per-component timing breakdown measured on a single NVIDIA A100 40GB GPU. The pipeline operates at stride-16, with CLIP extraction (0.14s/batch of 24). Dual-signal gating selects 102,066, 53,283, and 9,353 frames (69.7%, 76.5%, 65.1%) on XD, UCF, and MSAD for captioning (Qwen3-VL-8B, 8.19s/batch) and LLM scoring (Qwen2.5-7B, 0.64s/batch). At the video level, 15,968, 3,300, and 2,327 anchors pass the τf gate and undergo fusion via VideoLLaMA3-7B (0.84s) followed by LLM scoring (0.64s/batch). Table 4 compares SAAF against Unified [14]. By amortizing inference across total frames, SAAF reaches 44.95, 48.61, and 41.06 FPS on XD, UCF, and MSAD, delivers 2.43×, 2.63×, and 2.22× speedups over Unified 18.5 FPS. XD-Violence UCF-Crime

140

Number of Videos

a)

# 175

# Frames processed

B T(s)

Component Frame # 3

ShanghaiTech MSAD

120 100 80 60 40 20 0

0-10

10-20

20-30

30-40

40-50

50-60

60-70

Computation Reduction (%)

70-80

80-90

Figure 5. Distribution of per-video VLM call reduction relative to uniform stride-16 sampling across benchmark datasets. Higher reduction indicates fewer VLM calls within the stride-16 candidate pool. Adaptive frame selection achieves 20–60% reduction for the majority of videos while preserving detection accuracy.

Table 4. Throughput and accuracy across XD, UCF, and MSAD. Measured under identical settings.† LAVAD does not report MSAD. XD Method

FPS

AUC

UCF FPS

AUC

Table 6. Ablation on rolling window size W for drift signal.

MSAD FPS

Window Size (W )

AUC-ROC (%)

5 10 15 20

91.77 91.79 91.80 92.07

AUC

LAVAD † [38] 13.0 85.36 13.0 80.28 – – Unified [14] 18.5 91.34 18.5 84.28 18.5 85.90 SAAF (Ours) 44.95 92.07 48.61 82.59 41.06 86.70 Table 5. Backbone ablation for signal extraction on XD. All encoders achieve comparable performance. Drift window W = 20

Backbone CLIP ViT-B/32 DINOv3 ViT-B/14 DINOv2 ViT-B/14 SigLIP2 ViT-B/16

AUC-ROC (%) 92.07 91.65 91.78 91.61

6. Ablation Study Signal Attribution Distribution Figure 6 attributes each selected frame to δ-only, ρ-only, or both. Across all four benchmarks, δ-only accounts for 11–16% of selections and ρ-only for 14–24%, while 59–75% trigger both signals simultaneously. The non-trivial single-signal proportions (25–40% combined) represent frames a single-signal approach would miss, directly explaining the performance margin in Supplementary Sec. A.1 (Table 10).

ning W preceding frames. We evaluate window sizes W ∈ {5, 10, 15, 20} all others fixed. Table 6 reports the results. Performance improves slightly as window size increases. Smaller windows (W = 5) respond quickly to scene changes but produce noisier drift estimates, larger windows (W ≥ 20) accumulate sufficient temporal context to distinguish genuine drift from short-term fluctuations. We set W = 20 as the default. Stage-Wise Breakdown Analysis Table 7 reports cumulative AUC-ROC across pipeline stages. Each stage yields monotonic gains. λ-weighted video-level gating contributes +10.23% on XD and +11.57% on UCF, and refinement adds a further +6.32% on UCF and +5.04% on Shang. Table 7. Stage-wise ablation across benchmarks. AUC-ROC (%).

Stage

XD

UCF

MSAD

Shang

2. Frame-level 3. + λ-gating 4. + Refinement

78.30 88.53 92.07

64.70 76.27 82.59

72.13 82.36 86.70

71.38 75.50 80.54

Ablation of the Refinement Stage To isolate the contribution of conditional video-level fusion from Gaussian smoothing, we apply smoothing directly to Stage 2 outputs. Smoothing alone improves AUC-ROC on every benchmark, yet the full Stage 3 pipeline retains margins of 2.72, 9.29, 4.92, and 3.35 points on XD, UCF, MSAD, and Shang. The gap shows that video-level fusion supplies semantic evidence beyond what smoothing can provide. Table 8. Ablation of refinement components. Figure 6. Signal distribution (%) among selected stride-16 frames.

Configuration

Backbone Selection for Signal Extraction We evaluate four pretrained visual encoders for computing delta and drift signals. Table 5 reports AUC-ROC under identical pipeline settings. All backbones achieve comparable performance, with AUC-ROC spanning a narrow 0.46% range. CLIP marginally outperforms others, while Dinov2 ViTB/14 trails by only 0.29%. This shows that the dual-signal mechanism generalizes across embedding spaces, capturing anomaly-relevant temporal patterns regardless of the underlying encoder. We adopt CLIP ViT-B/32 as the default. Rolling Window Size for Drift Signal The drift signal computes deviation from a rolling temporal anchor span-

Stage 2 (no smoothing) Stage 2 + smoothing Stage 3 + smoothing (Stage 3)

XD-Viol.

UCF

MSAD Shang

78.30 89.35 92.07

64.70 73.30 82.59

72.13 81.78 86.70

71.38 77.19 80.54

7. Conclusion We presented SAAF, an efficient zero-shot VAD framework that resolves the tension between semantic capability and computational feasibility. Among training-free zeroshot methods, SAAF achieves SOTA on XD, MSAD, and Shang while reducing computation, suggesting that contentaware allocation guided by lightweight semantic signals is a promising direction for deploying VLMs at scale.

References [1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2 [2] Sunghyun Ahn, Youngwan Jo, Kijung Lee, Sein Kwon, Inpyo Hong, and Sanghyun Park. AnyAnomaly: zero-shot customizable video anomaly detection with LVLM. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3026–3035, 2026. 2, 3, 6 [3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob L Menick, Sebastian Borgeaud, Andy Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikoł aj Bińkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karén Simonyan. Flamingo: a visual language model for few-shot learning. In Advances in Neural Information Processing Systems, pages 23716–23736. Curran Associates, Inc., 2022. 2 [4] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 1 [5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-toend object detection with transformers. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, pages 213–229. Springer, 2020. 1 [6] Yingxian Chen, Zhengzhe Liu, Baoheng Zhang, Wilton W. T. Fok, Xiaojuan Qi, and Yik-Chung Wu. MGFN: magnitude-contrastive glance-and-focus network for weakly-supervised video anomaly detection. In ThirtySeventh AAAI Conference on Artificial Intelligence, AAAI 2023, Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence, IAAI 2023, Thirteenth Symposium on Educational Advances in Artificial Intelligence, EAAI 2023, Washington, DC, USA, February 7-14, 2023, pages 387–395. AAAI Press, 2023. 6 [7] Khaled Dawoud, Zaigham Zaheer, Mustaqeem Khan, Karthik Nandakumar, Abdulmotaleb Elsaddik, and Muhammad Haris Khan. FusedVision: A knowledge-infusing approach for practical anomaly detection in real-world surveillance videos. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 4036–4046, 2025. 1, 2, 6, 7 [8] Prabhu Prasad Dev, Raju Hazari, and Pranesh Das. Mcanet: Multimodal caption aware training-free video anomaly detection via large language model. In Pattern Recognition 27th International Conference, ICPR 2024, Kolkata, India, December 1-5, 2024, Proceedings, Part XXXII, pages 362– 379. Springer, 2024. 6 [9] Amir Ghodrati, Babak Ehteshami Bejnordi, and Amirhossein Habibian. Frameexit: Conditional early exiting for ef-

ficient video recognition. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 15608–15618. Computer Vision Foundation / IEEE, 2021. 3 [10] Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2 [11] Mahmudul Hasan, Jonghyun Choi, Jan Neumann, Amit K. Roy-Chowdhury, and Larry S. Davis. Learning temporal regularity in video sequences. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 733–742. IEEE Computer Society, 2016. 2 [12] Bruno Korbar, Du Tran, and Lorenzo Torresani. Scsampler: Sampling salient clips from video for efficient action recognition. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 6231–6241. IEEE, 2019. 3 [13] Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 19730– 19742. PMLR, 2023. 2 [14] Dongheng Lin, Mengxue Qu, Kunyang Han, Jianbo Jiao, Xiaojie Jin, and Yunchao Wei. A unified reasoning framework for holistic zero-shot video anomaly analysis. In Advances in Neural Information Processing Systems, pages 29713– 29744. Curran Associates, Inc., 2025. 2, 3, 6, 7, 8 [15] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 26286–26296. IEEE, 2024. 6 [16] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection - A new baseline. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 6536–6545. Computer Vision Foundation / IEEE Computer Society, 2018. 1, 2 [17] Weixin Luo, Wen Liu, and Shenghua Gao. A revisit of sparse coding based anomaly detection in stacked rnn framework. In Proceedings of the IEEE international conference on computer vision, pages 341–349, 2017. 6 [18] Jefferson Ryan Medel and Andreas Savakis. Anomaly detection in video using predictive convolutional long short-term memory networks. arXiv preprint arXiv:1612.00390, 2016. 2 [19] Yue Meng, Chung-Ching Lin, Rameswar Panda, Prasanna Sattigeri, Leonid Karlinsky, Aude Oliva, Kate Saenko, and Rogerio Feris. Ar-net: Adaptive frame resolution for efficient action recognition. In European conference on computer vision, pages 86–104. Springer, 2020. 3 [20] Jakub Micorek, Horst Possegger, Dominik Narnhofer, Horst Bischof, and Mateusz Kozinski. Mulde: Multiscale log-

density estimation via denoising score matching for video anomaly detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18868–18877, 2024. 2, 6 [21] Hyunjong Park, Jongyoun Noh, and Bumsub Ham. Learning memory-guided normality for anomaly detection. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2020, Seattle, WA, USA, June 1319, 2020, pages 14360–14369. Computer Vision Foundation / IEEE, 2020. 2 [22] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 8748– 8763. PMLR, 2021. 2, 4 [23] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 779–788, 2016. 1 [24] Yihua Shao, Haojin He, Sijie Li, Siyu Chen, Xinwei Long, Fanhu Zeng, Yuxuan Fan, Muyang Zhang, Ziyang Yan, Ao Ma, et al. Eventvad: Training-free event-aware video anomaly detection. In Proceedings of the 33rd ACM International Conference on Multimedia, pages 2586–2595, 2025. 3, 6 [25] Guodong Shen, Yuqi Ouyang, and Victor Sanchez. Video anomaly detection via prediction network with enhanced spatio-temporal memory exchange. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 3728–3732. IEEE, 2022. 2 [26] Waqas Sultani, Chen Chen, and Mubarak Shah. Real-world anomaly detection in surveillance videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6479–6488, 2018. 1, 6 [27] Yu Tian, Guansong Pang, Yuanhong Chen, Rajvinder Singh, Johan W Verjans, and Gustavo Carneiro. Weakly-supervised video anomaly detection with robust temporal feature magnitude learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4975–4986, 2021. 6 [28] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. LLaMA: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 2 [29] Jue Wang and Anoop Cherian. GODS: generalized one-class discriminative subspaces for anomaly detection. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 8200–8210. IEEE, 2019. 2, 6 [30] Peng Wu, Jing Liu, Yujia Shi, Yujia Sun, Fangtao Shao, Zhaoyang Wu, and Zhiwei Yang. Not only look, but also listen: Learning multimodal violence detection under weak

supervision. In European conference on computer vision, pages 322–339. Springer, 2020. 6 [31] Peng Wu, Xuerong Zhou, Guansong Pang, Yujia Sun, Jing Liu, Peng Wang, and Yanning Zhang. Open-vocabulary video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18297–18307, 2024. 6 [32] Peng Wu, Xuerong Zhou, Guansong Pang, Zhiwei Yang, Qingsen Yan, Peng Wang, and Yanning Zhang. Weakly supervised video anomaly detection and localization with spatio-temporal prompts. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 9301–9310, 2024. 2, 3, 6 [33] Zuxuan Wu, Caiming Xiong, Yu-Gang Jiang, and Larry S Davis. Liteeval: A coarse-to-fine framework for resource efficient video recognition. Advances in neural information processing systems, 32, 2019. 3 [34] Zuxuan Wu, Caiming Xiong, Chih-Yao Ma, Richard Socher, and Larry S Davis. Adaframe: Adaptive frame selection for fast video recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1278–1287, 2019. 3 [35] An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report. arXiv preprint arXiv:2505.09388, 2025. 1, 6 [36] Yuchen Yang, Kwonjoon Lee, Behzad Dariush, Yinzhi Cao, and Shao-Yuan Lo. Follow the rules: Reasoning for video anomaly detection with large language models. arXiv preprint arXiv:2407.10299, 2024. 6 [37] Muchao Ye, Weiyang Liu, and Pan He. VERA: explainable video anomaly detection via verbalized learning of visionlanguage models. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 8679–8688, 2025. 2, 3, 6 [38] Luca Zanella, Willi Menapace, Massimiliano Mancini, Yiming Wang, and Elisa Ricci. Harnessing large language models for training-free video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18527–18536, 2024. 2, 3, 6, 8 [39] Boqiang Zhang, Kehan Li, Zesen Cheng, Zhiqiang Hu, Yuqian Yuan, Guanzheng Chen, Sicong Leng, Yuming Jiang, Hang Zhang, Xin Li, et al. Videollama 3: Frontier multimodal foundation models for image and video understanding. arXiv preprint arXiv:2501.13106, 2025. 6 [40] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In Proceedings of the 2023 conference on empirical methods in natural language processing: system demonstrations, pages 543–553, 2023. 1, 2 [41] Huaxin Zhang, Xiaohao Xu, Xiang Wang, Jialong Zuo, Xiaonan Huang, Changxin Gao, Shanjun Zhang, Li Yu, and Nong Sang. Holmes-vau: Towards long-term video anomaly understanding at any granularity. In Proceedings of the computer vision and pattern recognition conference, pages 13843–13853, 2025. 3, 6 [42] Hang Zhou, Junqing Yu, and Wei Yang. Dual memory units with uncertainty regulation for weakly supervised video

anomaly detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3769–3777, 2023. 6 [43] Liyun Zhu, Lei Wang, Arjun Raj, Tom Gedeon, and Chen Chen. Advancing video anomaly detection: A concise review and a new dataset. Advances in Neural Information Processing Systems, 37:89943–89977, 2024. 6

Supplementary Material for Efficient Zero-Shot Video Anomaly Detection via Signal-Aware Adaptive Fusion (SAAF) Supplementary Roadmap This supplementary document provides extended experimental analysis, qualitative results, and implementation details that complement the main paper. We organize the material as follows:

Table 9. Event-level recall of Stage 1 selection. Unhittable events contain no stride-16 candidate; SAAF miss events had a candidate but no signal crossed threshold. Dataset Events Recall (%) Miss Unhittable SAAF miss MSAD UCF XD Shang

240 156 1210 194

98.3 92.3 88.9 77.8

4 12 134 43

1 0 0 22

3 12 134 21

• Section A Additional ablation study. • Section B Additional qualitative results through representative success and failure cases. • Section C Prompt templates used throughout the pipeline. • Section D Limitations. • Section E Broader impact considerations.

Table 10. Signal component ablation (Stage 1 AUC-ROC %). ORgated consistently outperforms single-signal variants.

A. Additional ablation study

Dataset

Delta

Drift

OR

AND

We conduct systematic ablations on all four benchmarks to evaluate the contribution of each SAAF component.

XD-Violence UCF-Crime MSAD ShanghaiTech

72.89 63.60 67.11 70.12

74.87 63.83 70.42 69.60

78.30 64.70 72.13 71.38

68.64 62.62 64.95 68.02

A.1. Event-Level Recall of Stage 1 Selection We measure how many true anomalies Stage 1 discards before scoring, independently of AUC-ROC. Using the dualsignal (δ ∨ ρ on CLIP, N = 1, W = 20, stride 16) and pervideo binary ground truth, we define an event as a contiguous run of anomalous frames. An event is hit when at least one selected frame, the only passed to the VLM, falls within it, and missed otherwise. We separate misses by cause. An event is unhittable when the stride-16 grid places no candidate frame inside it, so the selector never sees it; this affects any stride-16 method equally. A SAAF miss is when a candidate frame was available but neither δ nor ρ fired, which is the only case the selector is responsible for. Table 9 report the results. On MSAD, UCF, and XD, Stage 1 recovers 98.3%, 92.3%, and 88.9% of events, and nearly all misses are selector decisions rather than sampling blind spots. ShanghaiTech is the exception at 77.8%, but 22 of its 43 missed events are unhittable under stride 16, reflecting the dataset’s brief, gradually evolving anomalies rather than a thresholding failure, consistent with the coarse temporal resolution noted in the main paper.

A.2. Signal Component Ablation Table 10 isolates the contribution of each signal in the adaptive frame selection. We evaluate four configurations: δonly (frame-to-frame change), ρ-only (context deviation), OR-gated (either signal triggers selection), and AND-gated (both signals trigger).

The results validate three design choices. First, ORgated combination consistently achieves the highest performance across all datasets, confirming that delta and drift signals capture complementary anomaly patterns. Second, AND-gated combination yields the worst results, as requiring both signals to fire misses anomalies that manifest through only one signature. Third, the relative contribution of each signal varies by dataset: drift-only outperforms delta-only on XD and MSAD where gradual anomalies are prevalent, while performance is comparable on UCF and Shang. This variation reflects the differing characteristics of anomalies across benchmarks.

A.3. Selection Threshold (N ) The parameter N in the adaptive thresholding formula (t) (t) τx,t = µx + N · σx controls detection sensitivity by determining how many standard deviations above the running mean a signal must exceed to trigger frame selection. Table 11 reports both selection rate (relative to uniform stride-16 sampling) and AUC-ROC across N ∈ {0.5, 1.0, 1.5, 2.0}. Lower N produces more permissive thresholds and higher selection rates, while higher N restricts selection to the most extreme deviations. The results reveal a coverage accuracy trade-off inherent to adaptive sampling. At N = 0.5, permissive thresholds achieve the highest AUC-ROC across all datasets but select 76–84% of

candidate frames, reducing the computational savings that motivate adaptive selection. At N = 2, aggressive filtering reduces selection to 43–59% but misses informative frames, with XD suffering the largest degradation relative to the default (−9.70 points, from 78.30 to 68.60). We adopt N = 1 as the default, achieving a practical balance: selection rates of 65–77% preserve the majority of anomalyrelevant frames while still providing efficiency gains over uniform sampling.

sd

XD-Viol.

UCF

MSAD

Shang

0 0.1 0.2 0.3

64.90 76.26 78.98 78.30

55.44 61.52 64.90 64.70

67.50 72.74 72.13 72.13

63.35 68.15 71.38 71.38

benchmarks while encoding a conservative normality prior that avoids over-penalizing unselected frames.

A.4. Default Fill Value (sd ) Frames that are not selected by the dual-signal mechanism receive a default score sd encoding a normality prior. Table 12 evaluates the sensitivity of Stage 1 performance to sd ∈ {0, 0.1, 0.2, 0.3}. Setting sd = 0 yields substantially degraded performance across all datasets, as unselected normal frames receive scores identical to confident normality predictions, compressing the score distribution and harming discriminability. Performance improves markedly with sd ≥ 0.1, which provides separation between unselected frames and explicitly scored normal content. The optimal value varies slightly by dataset: sd = 0.2 achieves the best results on XD-Violence, UCF-Crime, and ShanghaiTech, while MSAD peaks at sd = 0.1. We adopt sd = 0.3 as the default, which performs within 0.7% of optimal across all Table 11. Selection threshold ablation. Stage 1 AUC-ROC (%) with selection rate relative to uniform stride-16 sampling. Lower N yields higher selection rates but diminished efficiency gains.

Dataset

N

Sel. Rate (%)

AUC-ROC

XD-Violence

0.5 1.0 1.5 2.0

79.80 69.70 58.50 42.70

80.74 78.30 74.66 68.60

UCF-Crime

0.5 1.0 1.5 2.0

83.90 76.50 67.50 45.70

64.87 64.70 64.25 59.18

MSAD

0.5 1.0 1.5 2.0

76.40 65.10 61.10 55.40

75.23 72.13 72.37 71.61

0.5 1.0 1.5 2.0

82.10 72.40 65.70 59.40

73.46 71.38 70.86 69.94

ShanghaiTech

Table 12. Default fill value (sd ) ablation. Stage 1 AUC-ROC (%).

A.5. Fusion Threshold Sensitivity (τf ). We evaluate the effect of the fusion threshold τf on detection performance and computational cost across all four benchmarks. Table 13 reports AUC-ROC and the video-level invocation rate, the fraction of frames triggering cascaded VLM inference, for τf ∈ {0.4, 0.5, 0.6, 0.7, 0.8, 0.9}. Performance degrades gracefully as τf increases. On XD-Violence, AUC-ROC decreases by only 0.86% while the invocation rate drops from 10.9% to 4.5%, a 2.4× reduction in VLM calls. UCF-Crime and MSAD remain within 0.7%, and ShanghaiTech within 0.3%, across the full sweep. The invocation rate on ShanghaiTech collapses at τf ≥ 0.8 (from 19.5% to 0.1%), as the video-level model assigns low confidence scores to its subtle anomalies, reducing SAAF to frame-level scoring. We adopt τf = 0.4 as the default to maximize recall, but τf can be raised for substantial efficiency gains at negligible accuracy cost, making SAAF adaptable to varying deployment budgets. Table 13. Effect of fusion threshold τf on AUC-ROC (%) and invocation rate (%). Higher τf reduces VLM calls with minimal performance degradation. XD-Violence

UCF-Crime

τf

AUC

Inv.%

AUC

Inv.%

AUC

MSAD Inv.%

ShanghaiTech AUC

Inv.%

0.4 0.5 0.6 0.7 0.8 0.9

88.53 88.00 88.00 87.87 87.81 87.67

10.90 10.5 10.5 8.5 7.6 4.5

76.27 75.86 75.86 75.80 75.85 75.70

4.7 3.4 3.4 1.8 1.5 0.7

82.36 82.08 82.07 82.01 81.88 81.71

16.2 16.2 16.1 15.2 12.1 9.2

75.50 75.35 75.35 75.61 75.47 75.41

22.5 22.5 22.5 19.5 0.1 0.0

A.6. Signal-Aware Fusion Weights (λδ , λρ ) The signal-aware fusion mechanism assigns weights λδ and λρ to frame-level scores based on which signal triggered selection: δ-triggered frames weight frame-level scores by λδ , while ρ-triggered frames use λρ . When both signals fire, we apply λboth = 0.5. Table 14 evaluates four weight configurations across all datasets.

Table 14. Signal-aware fusion weight ablation. Stage 2 AUC-ROC (%) under different (λδ , λρ ) configurations.

Config

λδ

λρ

XD

UCF

MSAD

Shang

firm moderate default reversed

0.8 0.7 0.6 0.2

0.2 0.3 0.4 0.8

88.07 88.38 88.53 88.14

75.98 76.19 76.27 76.08

82.08 82.30 82.36 81.73

75.35 75.31 75.50 76.43

We adopt (λδ , λρ ) = (0.6, 0.4) as our default. δ activations mark abrupt visual transitions whose evidence is concentrated in the triggering frame, so frame-level scoring dominates the fused decision. ρ activations mark persistent or gradually evolving states whose interpretation benefits from temporal context, so the video-level score receives higher weight. This weighting aligns fusion with the semantics of the triggering signal. Across XD-Violence, UCFCrime, and MSAD, performance varies by less than 0.4 points across configurations, indicating the stage is robust to this choice. ShanghaiTech is the exception, where a ρdominant weighting (0.2, 0.8) yields a 0.93-point gain, reflecting the dataset’s bias toward gradually evolving anomalies in a single scene. We retain (0.6, 0.4) as the default to avoid dataset-specific tuning.

A.7. Additional Per-Category Analysis Figure 7 presents the per-category AUC-ROC breakdown across the three SAAF pipeline stages on MSAD. Strong performance categories. SAAF achieves nearperfect detection (>95% AUC-ROC after refinement) on high-motion sudden events such as Fighting (99.8%), Explosion (98.5%), Robbery (98.0%), Shooting (97.8%), Assault (96.6%), and Object falling (95.7%). These categories exhibit strong visual signatures such as rapid motion, dramatic scene changes, or impacts that both the δ signal and frame-level VLM scoring capture reliably. Challenging categories. Gradual-onset and visually subtle anomalies remain more difficult. Vandalism (85.7%) involves fine-grained actions that may not produce distinctive δ or ρ signals until the act is underway. Fire (89.2%) is limited by slow visual progression that the drift signal must accumulate over time. Stage-wise improvement patterns. The multi-stage architecture provides consistent gains across all categories. Video-level gating (Stage 2) delivers the largest improvements on categories requiring temporal context, with Object falling improving by 21.5% and People falling by 15.1%. Temporal refinement (Stage 3) further boosts performance by 2–10% across categories. These patterns indicate that sudden events benefit from accurate frame-level detection, while ambiguous or gradual anomalies prefer video-level reasoning.

B. Additional qualitative results This section presents qualitative analysis of SAAF’s behavior through representative success and failure cases, illustrating both the framework’s strengths and its current limitations.

B.1. Qualitative Success Cases We present representative examples where SAAF successfully detects anomalies while achieving substantial computational savings. Figure 8 illustrates four spanning anomalous events and normal content across different datasets. Efficient detection of sudden events. Figure 8(a) demonstrates SAAF operating effectively on a dashcamcaptured road accident from XD-Violence. The dual-signal mechanism correctly concentrates activations around the anomalous region (frames 650–750, shaded), where a sudden explosion and car accident occurs. The δ signal responds sharply to the visual transition as the explosion ignites, while both signals appropriately skip the preceding normal driving footage (frames 200–630) where the scene remains static. The VLM produces a high anomaly score (0.65) precisely within the event region, providing clear separation from normal frames. The computational efficiency is substantial, out of 63 candidate frames, only 15 are selected for stage1 and stage 2, a selection ratio of 23.8%. This translates to 76.2% computational savings compared to uniform stride-16 sampling, as the expensive VLM pipeline is invoked only when the dual-signal mechanism detects meaningful content. This case exemplifies SAAF’s design goal, directing expensive computation toward anomaly-relevant content while aggressively skipping redundant frames, achieving both high detection accuracy and substantial efficiency gains. Detecting violations in pedestrian zones. Figure 8(b) shows a cycling anomaly from ShanghaiTech, where a bicycle enters a pedestrian-only area. The dual-signal mechanism responds robustly to the anomalous region (frames 160–220, shaded), with both δ and ρ signals firing as the cyclist appears and moves through the scene. The Stage 2 scoring produces high anomaly scores (0.8–0.85) precisely within the event window, correctly identifying the violation. Notably, the signals also activate during the post-anomaly period (frames 250–350) as pedestrian activity increases, though scores appropriately decrease (0.3–0.55) reflecting absent anomaly presence. Out of 38 stride-16 candidate frames, 22 are selected for VLM analysis, a selection ratio of 57.9%, yielding 42.1% computational savings compared to uniform sampling. Efficient processing of normal content. Figure 8(c) demonstrates SAAF’s behavior on a normal video from UCF-Crime containing no anomalous events (all-zero ground truth). The scene shows a static industrial environment with minimal activity across 1,600+ frames. The dual-

Figure 7. MSAD (11 anomaly categories)

(a) Success case: Road accident captured from dashcam (XD-Violence). SAAF concentrates frame selection around the anomalous event, achieving 76.2% computational savings with high detection accuracy.

(b) Success case: Cycling anomaly in pedestrian area (ShanghaiTech). SAAF accurately detects the anomalous activity while skipping normal pedestrian scenes, achieving 42.1% computational savings.

(c) Success case: Normal video (UCF-Crime). SAAF correctly identifies this as a non-anomalous scene, maintaining low scores while achieving 89.4% computational savings through smart frame selection.

(d) Success case: Normal video from metro station (XD-Violence). Despite continuous pedestrian activity, SAAF maintains low anomaly scores and achieves 84.5% computational savings.

Figure 8. Representative success cases demonstrating SAAF’s detection accuracy and computational efficiency across anomalous (a, b) and normal (c, d) videos.

signal mechanism correctly recognizes the lack of meaningful content: activations occur only during brief initial dynamics (frames 0–200) and minor scene changes near the end (frames 1,500–1,600), while the vast majority of frames are appropriately skipped.

The efficiency gains are substantial, only 12 out of 113 stride-16 candidate frames are selected for VLM analysis, a selection ratio of 10.6%, yielding 89.4% computational savings compared to uniform sampling. Throughout the video, anomaly scores remain consistently low, correctly reflecting

(a) Failure case: Robbery 1 (MSAD). The anomaly begins at frame 0, contaminating baseline statistics during warmup and preventing proper signal activation.

(b) Failure case: RoadAccidents011 (UCF-Crime). Intense, fast-moving traffic triggers dense signal activations throughout the video. The limited computation savings are attributed to the fast-changing scene.

(c) Failure case: Arson007 (UCF-Crime). The dual-signal mechanism correctly identifies the anomalous region, but the VLM fails to recognize the subtle fire progression.

(d) Failure case: Fire 12 (MSAD). The dual-signal mechanism correctly concentrates selections in the anomalous region, but the VLM fails to distinguish subtle fire event.

Figure 9. Representative failure cases illustrating SAAF’s limitations.

the absence of anomalous content. Handling dynamic normal scenes. Figure 8(d) presents a challenging normal video from XD-Violence depicting a busy metro station with continuous pedestrian flow. Unlike the static industrial scene in Figure 8(c), this environment contains persistent motion from commuters walking, entering, and exiting the frame. The dual-signal mechanism demonstrates appropriate calibration: activations concentrate during initial scene establishment (frames 0–200) and brief periods of elevated activity (frames 1,200–1,400), while the extended middle portion (frames 200–1,200) is correctly skipped despite ongoing pedestrian movement. Crucially, Stage 2 produces consistently low scores (0.02–0.08) for all selected frames, correctly classifying the normal commuter behavior. The computational efficiency remains high even in this dynamic setting: only 16 out of 103 stride-16 candidate frames require Stage 1 & Stage 2 analysis, a selection ratio of 15.5%, yielding 84.5% computational savings. This case demonstrates that SAAF’s efficiency gains extend beyond static scenes to dynamic environments with regular

activity patterns, invoking VLM inference only when scene dynamics deviate from established baselines rather than responding to all motion indiscriminately.

B.2. Failure Case Analysis We analyze representative failure cases to understand SAAF’s limitations and identify directions for future improvement. Figure 9 presents four scenarios that expose distinct failure modes. Baseline contamination from early-onset anomalies. Figure 9(a) reveals a structural limitation of SAAF’s adaptive thresholding mechanism. The robbery event begins immediately at frame 0, meaning the anomalous content is present during the warmup period when baseline statistics (µ and σ) are initialized. As a result, the EMA-based thresholds adapt to the anomaly as if it were normal behavior, contaminating the baseline distribution. The consequence is visible in the signal activation pattern, despite the entire shaded region (frames 0–200) being anomalous, the dualsignal mechanism produces sparse activations because the δ and ρ values do not exceed thresholds calibrated on anoma-

lous content. Notably, it assigns high scores (0.7–0.9) for the frames that are selected, demonstrating that the scoring stage functions correctly, the failure lies entirely in frame selection due to corrupted baseline initialization. This case exposes an assumption of our approach, the warmup period must contain representative normal content to establish meaningful thresholds. Videos where anomalies begin immediately at the early frames violate this assumption, causing the adaptive mechanism to normalize abnormal patterns. Brief anomalies in high-motion environments. Figure 9(b) shows a road accident from UCF. The scene depicts a busy intersection with continuous vehicle and pedestrian movement, causing the dual-signal mechanism to activate densely throughout the entire video, both within and outside the brief anomalous region (frames 250–300, shaded). The accident itself, a brief collision lasting only 50 frames, produces visual cues similar to vehicles braking or turning abruptly. This case illustrates SAAF’s difficulty with transient anomalies embedded in dynamic scenes, where VLM ambiguity regarding subtle collision cues results in poor separation between anomalous and normal content. Subtle anomalies beyond VLM recognition. Figure 9(c) shows an arson case from UCF-Crime where δ and ρ fire correctly within the anomalous region (frames 2,500– 5,500, shaded) yet anomaly scores remain low throughout the event. The night-vision footage shows gradual fire progression with faint smoke and small flames against a cluttered background of Christmas lights, visual cues that fall below the VLM’s detection capability. When anomalies exhibit signatures that challenge current VLMs, accurate selection alone cannot compensate. Visually subtle fire in industrial setting. A similar pattern emerges in Figure 9(d), which shows a fire event from MSAD occurring in an industrial environment with HVAC equipment. The signal activation pattern confirms that SAAF’s frame selection operates correctly: the dual-signal mechanism appropriately concentrates activations within the anomalous region (shaded) where fire develops. However, the anomaly scores remain low throughout the event, failing to distinguish the fire from normal frames.

C.1. Frame-Level Visual Captioning Prompt The captioning prompt pcap instructs Qwen3-VL-8B to generate comprehensive scene descriptions with emphasis on relevant details. The prompt is structured to extract information across multiple dimensions relevant to anomaly detection.

Frame-Level Captioning Prompt (pcap ) Analyze this footage frame and provide scene description: - All people present: their count, positions, activities, and interactions - Any unusual behaviors or events - Objects of interest: bags, weapons, vehicles, or suspicious items - Environmental context: location type (street, pedestrian area, parking lot, etc. - Vehicles in pedestrian areas: note any cars, bicycles, skateboards, or carts in pedestrian-only zones - Any safety or security concerns

The structured format encourages the VLM to attend to specific visual elements that correlate with anomalous activity: person count and behavior, object presence, environmental context, and potential safety concerns. This decomposition ensures that downstream reasoning receives rich descriptions rather than generic scene summaries.

C.2. Anomaly Scoring Prompt

C. Prompts and Templates This section provides the prompt templates used throughout the SAAF pipeline. All prompts are designed to elicit structured, task-relevant outputs from the respective models while maintaining consistency across inference stages. We present three core prompts: the frame-level captioning prompt (pcap ) for visual description, the anomaly scoring prompt (pscore ) for classification reasoning, and the videolevel temporal prompt (pvid ) for multi-frame analysis. We conclude with examples demonstrating the inference chain.

The scoring prompt pscore provides Qwen2.5-7B-Instruct with explicit criteria for anomaly classification, used for both frame-level and video-level. The prompt enumerates concrete indicators for both anomalous and normal scenarios. The enumeration of specific anomaly categories aligns with the types present in the evaluation benchmarks (XD, UCF, MSAD, ShanghaiTech), while the normal indicators help calibrate the model’s baseline expectations. The LLM outputs a scalar score s ∈ [0, 1] reflecting the assessed likelihood of anomalous content.

Anomaly Classification System Prompt (pscore ) You are an expert AI system specialized in video anomaly detection. Analyze scene descriptions and classify them as NORMAL or ANOMALY. ANOMALY INDICATORS: - Violence: fighting, assault, weapons, threatening behavior - Explosions: fire, smoke, blast effects, debris, damaged structures - Accidents: car collisions, crashes, overturned vehicles - Theft: shoplifting, robbery, stealing, grabbing items and fleeing - Medical emergencies: injuries, falls, people on ground needing help - Suspicious behavior: inappropriate running, chasing, panic, fleeing - Property damage: vandalism, destruction, breaking - Abandoned objects: unattended bags, suspicious packages - Crowd anomalies: panic, stampede, unusual mass movements - Fire or smoke visible NORMAL INDICATORS: - Routine activities: walking, standing, sitting, talking normally - Shopping, browsing, work, recreational activities - Vehicles in designated areas: roads, parking lots, bike lanes - Regular traffic flow - Appropriate social interactions - quiet scenes with no damage or suspicious objects -Respond ONLY with a JSON object containing the anomaly score: "anomaly score": 0.0-1.0

Temporal Multi-Frame Captioning Prompt (pvid ) Describe the actions, movements, and events visible in this sequence. Specifically identify any suspicious, unusual, or anomalous behavior such as: theft, shoplifting, violence, fighting, accidents, burglary, or other criminal activities.

This prompt is deliberately more concise than the framelevel captioning prompt, as the video model receives multiple frames providing temporal context.The emphasis on actions and movements leverages the video model’s capacity to reason about dynamics across frames, complementing frame-level static analysis.

C.4. Example Pipeline Outputs To illustrate the complete inference chain, we provide representative examples showing the progression from visual input through captioning to final anomaly scoring. Example 1: Anomalous Scene (Explosion)

Delta-Triggered Frame (Sudden Event) Generated Caption: “The scene shows a large explosion occurring in an urban street. Thick black smoke and orange flames are visible rising from the center of the frame. Debris is scattered across the road. Several people are running away from the blast site in apparent panic. Damaged vehicles are visible near the explosion epicenter. The environmental context suggests a commercial district with storefronts visible in the background.” Triggering Signal: Delta (δt = 0.42, threshold τδ,t = 0.18) Anomaly Score: sframe = 0.95 Video-Level Gating: sframe = 0.95 > τf = 0.4 → invoke video-level analysis Video-Level Score: svid = 0.91

C.3. Video-Level Temporal Captioning Prompt

Fusion Weight: λδ = 0.6 (delta-triggered → frame-level dominant) Fused Score: sfused = λδ · sframe + (1 − λδ ) · svid = 0.6 × 0.95 + 0.4 × 0.91 = 0.934

For frames that exceed the threshold τf , we extract a temporal clip and process it through VideoLLaMA3-7B using the following prompt pvid :

Example 2: Anomalous Scene (Static Anomaly)

Drift-Triggered Frame (Persistent Anomaly) Generated Caption: “A person is lying motionless on the ground in a pedestrian walkway. The individual appears to be in a prone position with no visible movement. Several pedestrians are walking past without stopping. An unattended bag is visible next to the person. The scene suggests a possible medical emergency or collapse.” Triggering Signal: Drift (ρt = 0.31, threshold τρ,t = 0.12) Anomaly Score: sframe = 0.78 Video-Level Gating: sframe = 0.78 > τf = 0.4 → invoke video-level analysis Video-Level Score: svid = 0.85 (confirms persistent anomaly) Fusion Weight: λρ = 0.4 (drift-triggered → video-level dominant) Fused Score: sfused = λρ · sframe + (1 − λρ ) · svid = 0.4 × 0.78 + 0.6 × 0.85 = 0.822

Example 3: Normal Scene (Unselected Frame) Unselected Frame (Below Signal Thresholds) Delta Signal: δt = 0.05 < τδ,t = 0.15 → not triggered Drift Signal: ρt = 0.08 < τρ,t = 0.11 → not triggered Frame Selection: I(t) = Iδ (t) ∨ Iρ (t) = 0 → frame not selected VLM Analysis: Skipped (no captioning or scoring performed) Final Score: st = sd = 0.3 (default normality prior)

These examples illustrate the complete SAAF pipeline: frames are selected only when delta or drift signals exceed their adaptive thresholds, video-level analysis is invoked only when the frame-level score exceeds τf = 0.4, and unselected frames receive the default score sd

D. Limitations SAAF’s adaptive thresholding assumes the warmup period contains representative normal content to establish baseline statistics. When anomalies begin immediately or dominate early frames, the EMA-based thresholds adapt to abnormal patterns as if they were normal, suppressing subsequent signal activations. This baseline contamination effect is evident in videos where anomalous events start at frame zero. The dual-signal mechanism operates on stride-16 sampled frames, which limits temporal resolution for brief anomalies. On ShanghaiTech, where events typically span 40–60 frames, this coarse sampling provides only 2–4 signal observations per anomaly, reducing detection reliability

compared to longer-duration events. Finally, SAAF’s accuracy ultimately depends on VLM reasoning capability. When anomalies exhibit subtle visual signatures, such as gradual fire progression in low-light conditions or fine-grained actions, correct frame selection cannot compensate for limitations in downstream semantic understanding. This coupling means that advances in VLM capabilities would directly benefit SAAF without architectural changes.

E. Broader Impact SAAF’s zero-shot design avoids encoding dataset-specific biases into learned parameters, though biases in underlying VLMs may still influence scoring. From an environmental perspective, SAAF’s sampling reduces substantial computational cost compared to uniform processing, lowering the energy footprint of large-scale video analysis. As VLM deployment scales across infrastructure worldwide, such efficiency gains become increasingly consequential for sustainable AI.

Related documents

• Efficient Zero-Shot Video Anomaly Detection via Signal-Aware Adaptive Fusion (SAAF)#1447853
• NOVA: Normal-Side Modeling for Training-Free Zero-Shot Video Anomaly Detection#973941
• A VLM Answer Is Not an Anomaly Score: Rank Compression Across Image and Video Anomaly Detection#987795
• Test-time Reinforcement Learning for Anomalous Video Understanding#1062002
• Anomaly-Guided Observation and Two-Level Semantic Feedback for Training-Free Video Anomaly Analysis#1327914