PoinCLIP-VAD: a hyperbolic cross-modal fusion framework for video anomaly detection Authors: Debi Prasad Senapati, Santosh Kumar Pani, Santos Kumar Baliarsingh, Prabhu Prasad Dev, Ramasubbareddy Somula Published: 2026 ABSTRACT AbstractWeakly supervised video anomaly detection (WSVAD) is fundamentally constrained by the absence of frame-level annotations, which leads to noisy instance selection in Multiple Instance Learning (MIL) and weak correspondence between temporal video segments and semantic descriptions. Vision-language models address this by enabling cross-modal alignment between visual features and textual labels, but when these representations are learned in Euclidean space, they struggle to capture subtle semantic variations and often produce ambiguous instance ranking under weak supervision. To address this limitation, we propose PoinCLIP-VAD, a vision-language framework that performs cross-modal fusion in hyperbolic space. The model embeds visual and textual features into a shared Poincaré ball geometry, where non-linear distance scaling provides a more expressive representation of latent semantic relationships induced by cross-modal interactions, without relying on predefined hierarchical structures. This geometry-consistent formulation enables more reliable similarity estimation and better preserves distinctions between normal and anomalous patterns. The framework adopts a dual-block architecture consisting of a classification block for coarse anomaly scoring and a video-text alignment block for fine-grained correspondence using negative Poincaré distance. Extensive experiments on benchmark datasets demonstrate that PoinCLIP-VAD achieves an AUC of 90.62% on UCF-Crime and an AP of 86.93% on XD-Violence, confirming improved anomaly discrimination and more consistent cross-modal alignment under weak supervision.Keywords: Video anomaly detection, Multiple instance learning, Weakly supervised learning, Cross-modal interaction, Vision–language modeling, Hyperbolic manifold, Poincaré distanceSubject terms: Engineering, Mathematics and computing Introduction Video anomaly detection (VAD) focuses on recognizing unusual activities or behaviors in surveillance videos, making it a vital and challenging aspect of video understanding in computer vision. Recently, VAD has gained more attention from both academia and industry due to its wide range of practical applications, including autonomous vehicle systems1, accident anticipation2, and security surveillance3. Initial research relied primarily on semi-supervised approaches where models were trained only on normal videos to learn normal behavioral patterns. Any behavior deviating from these learned norms during testing was considered abnormal. Such methods do not rely on a large number of labeled datasets, which makes them particularly feasible in real-world scenarios where anomalies are rare or difficult to define. These approaches tend to generalize well across diverse environments. However, due to the absence of explicit modeling of abnormal patterns, they may misclassify complex or previously unseen events, leading to false positives or missed detections. To address this limitation, recent studies adopt weakly supervised learning frameworks, formulating the problem as a Multiple Instance Learning (MIL) task4–7. In this setting, each video is treated as a bag of temporal instances, and predictions are derived from video-level annotations rather than frame-level supervision8. Following the MIL paradigm, several weakly supervised video anomaly detection approaches have been proposed to localize abnormal segments using only video-level labels. Zhu et al.9 introduced a temporal augmented network to capture motion-aware features via attention mechanisms. Li et al.10 further employed transformer-based architectures with multi-sequence learning to mitigate instance selection errors. Recent graph-based methods leverage message passing to model temporal dependencies between video snippets, demonstrating the ability to capture structural relations over time. However, these approaches typically operate within Euclidean embedding spaces, which treat distances uniformly and may limit the model’s ability to distinguish subtle deviations between normal and anomalous events. Weakly supervised video anomaly detection is inherently challenged by noisy instance selection in MIL, limited separability between normal and anomalous segments under weak supervision, and difficulties in achieving reliable cross-modal alignment between visual snippets and semantic descriptions. Although dataset annotations are flat, no explicit hierarchical structure is defined in the data, and any structural organization is implicitly learned within the embedding space through vision-language alignment. In parallel, large-scale vision-language models such as CLIP11 have advanced multimodal understanding through contrastive learning on image-text pairs. CLIP aligns visual and textual representations in a shared Euclidean space and demonstrates strong generalization capabilities. However, its reliance on Euclidean cosine similarity may limit its ability to capture fine-grained semantic variations between modalities. To address this limitation, PoinCLIP12 reformulates CLIP within a hyperbolic space by projecting representations into the Poincaré ball manifold and computing similarity using negative Poincaré distance. This formulation provides a more expressive geometric space for modeling complex semantic relationships. Motivated by these observations, this paper extends curvature-aware representation learning to weakly supervised video anomaly detection and introduces PoinCLIP-VAD, a hyperbolic vision-language framework for anomaly reasoning through cross-modal fusion. Unlike conventional Euclidean approaches, PoinCLIP-VAD embeds both visual and textual features within a shared hyperbolic space using exponential mapping, enabling more effective representation learning across modalities. The framework adopts a dual-block architecture consisting of a classification block for estimating coarse anomaly confidence and a video-text alignment block for enforcing fine-grained correspondence between frame-level visual features and textual label embeddings using negative Poincaré distance. Through this geometry-consistent design, PoinCLIP-VAD facilitates robust and interpretable anomaly detection under weak supervision. The key contributions of this work are summarized as follows: We propose PoinCLIP-VAD, a curvature-aware vision–language framework that integrates hyperbolic geometry into weakly supervised video anomaly detection, embedding visual and textual representations in the Poincaré ball manifold to enable geometry-consistent multimodal learning and more stable anomaly detection under weak supervision. We propose PoinCLIP-VAD, a curvature-aware vision–language framework that integrates hyperbolic geometry into weakly supervised video anomaly detection, embedding visual and textual representations in the Poincaré ball manifold to enable geometry-consistent multimodal learning and more stable anomaly detection under weak supervision. We design a dual-block architecture that jointly performs coarse anomaly classification and fine-grained video-text alignment within a hyperbolic space, where similarity is computed using negative Poincaré distance to ensure consistency with the underlying geometry. We design a dual-block architecture that jointly performs coarse anomaly classification and fine-grained video-text alignment within a hyperbolic space, where similarity is computed using negative Poincaré distance to ensure consistency with the underlying geometry. We develop a unified optimization framework that combines binary classification, cross-modal alignment, and contrastive objectives, enabling effective and stable training under weak supervision. We develop a unified optimization framework that combines binary classification, cross-modal alignment, and contrastive objectives, enabling effective and stable training under weak supervision. Literature survey Unsupervised video anomaly detection In the early stages, video anomaly detection was commonly framed as a one-class classification problem, where models were trained only on normal instances and later tested to recognize deviations from learned patterns13. Previous studies on unsupervised video anomaly detection mainly relied on handcrafted feature extraction to model normal patterns and identify deviations as anomalies. Trajectory-based approaches learned regular motion behaviors and detected abnormal trajectories that significantly deviated from the learned distributions14. Other techniques utilized dense motion and appearance cues, such as histograms of oriented gradients (HOG), histograms of optical flow (HOF), and spatiotemporal gradients within local cuboids, to construct statistical representations of normal activities15. Wang et al.16 modeled both motion and appearance in crowded scenes using a mixture of dynamic textures (MDT) and treated low-probability events under this model as anomalies. Similarly, cell-based frameworks extracted local descriptors of speed, size, and texture within each frame region and applied probabilistic models such as kernel density estimation or codebook learning to distinguish normal from abnormal motion patterns17. While these handcrafted feature-based methods offered interpretability and required no labeled data, they often suffered from sensitivity to illumination changes, occlusions, and camera motion, limiting their generalization to complex real-world surveillance environments. With the rapid growth of deep learning, the paradigm of anomaly detection shifted toward representation learning using neural networks. Zaheer et al.18 introduced a fully unsupervised framework for video anomaly detection called Generative Cooperative Learning (GCL). The approach jointly trains a generator and a discriminator in a cooperative manner using unlabeled videos, where each network alternately produces pseudo labels to guide the other. The generator reconstructs dominant normal patterns while distorting high-confidence anomalies through negative learning, and the discriminator learns to classify anomalous instances using these pseudo labels. This mutual learning strategy enables the location of anomalies without manual supervision, effectively leveraging the natural imbalance between normal and anomalous events in real-world surveillance videos. Al-lahham et al.19 proposed C2FPL (Coarse-to-Fine Pseudo-Labeling), a fully unsupervised framework for video anomaly detection that eliminates the need for manual annotations. The method introduces a two-stage pseudo-label generation process–coarse video-level labeling using hierarchical divisive clustering and fine segment-level refinement via statistical hypothesis testing. These generated pseudo labels are then used to train a lightweight anomaly detector in a supervised manner, enabling accurate and efficient frame-level anomaly identification across large-scale surveillance datasets. Al-lahham et al.20 proposed CLAP (Collaborative Learning of Anomalies with Privacy), a fully unsupervised video anomaly detection framework that enables distributed and privacy-preserving training without sharing raw data. The method introduces a collaborative architecture where multiple participants independently train local models using pseudo-labeled video segments generated via Von Neumann entropy and Gaussian Mixture Modeling. Through server knowledge accumulation and local feedback mechanisms, CLAP aggregates global model parameters using federated averaging while refining pseudo labels iteratively, achieving anomaly localization in large-scale surveillance scenarios under a federated learning configuration. Sun et al.21 introduced Dual GroupGAN, an unsupervised video anomaly detection framework that addresses overgeneralization and noise sensitivity in reconstruction- and prediction-based methods. The model employs a four-competitor (2V2) design incorporating two generators: SE-U-Net for prediction and SE-VAE for reconstruction, both enhanced with Squeeze-and-Excitation (SE) channel attention modules. A weighted fusion strategy combines prediction and reconstruction errors to balance spatial-temporal feature learning, enabling more robust and discriminative anomaly detection across diverse surveillance scenarios. Cao et al.22 proposed SAMN, a reverse distillation-based framework that integrates scale-aware modulation and a mixture of experts to adaptively capture multi-scale features and reconstruct normal patterns for precise medical anomaly detection. Kim et al.23 proposed a two-stream autoencoder with memory networks that jointly learn spatial features in Euclidean space and hierarchical temporal dynamics in constant curvature manifold space, enabling enhanced reconstruction error amplification for accurate video anomaly detection. Wu et al.24 proposed MCMP, a multi-cluster memory prototype framework that leverages contrastive clustering, global memory prototypes, and Gaussian filtered feature transformation to capture fine and coarse-grained variations in normal video patterns and improve anomaly discrimination through reconstruction error. Liu et al.25 propose MS2NL, a dual decoder framework with multi-scale spatial fusion, dynamic temporal aggregation, and a memory network to learn prototypical normal patterns and detect anomalies via reconstruction and motion prediction errors across scales. Lyu et al.26 propose a Bidirectional Skip-frame Prediction (BiSP) framework with variance channel attention and context spatial attention to enhance unsupervised video anomaly detection through bidirectional temporal prediction and intra-domain disparity modeling. Weakly supervised video anomaly detection Weakly Supervised Video Anomaly Detection (WSVAD) has attracted growing interest as it requires only video-level annotations, eliminating the need for costly frame-level labeling. Karim et al.27 introduced REWARD, a real-time end-to-end framework for weakly supervised video anomaly detection that trains large video models directly on raw inputs. It replaces metric-learning losses with a self-supervised kNN-based pseudo-labeling mechanism to enable efficient frame-level training. The approach eliminates feature aggregation, simplifying the wVAD pipeline while supporting real-time inference. Fan et al.28 proposed a Snippet Anomalous Attention framework that integrated a temporal embedding unit to capture both local and global temporal dependencies, along with an anomalous attention module that emphasized informative snippets without depending on pseudo labels. A multi-branch supervision mechanism further refined the learning process by suppressing dominant regions and improving anomaly completeness, leading to more accurate anomaly localization. Ullah et al.29 proposed a hybrid CNN-TCN model that combined convolutional feature extraction with a multihead attention-based temporal convolutional network to effectively capture short-term and long-term dependencies. By employing dilated convolutions, residual links, and multihead temporal weighting, their framework extracted highly discriminative spatiotemporal representations, resulting in improved recognition and classification of abnormal events in surveillance videos. Abiew et al.30 proposed HiTESS, a hierarchical temporal sequence segmentation framework that leveraged multi-scale temporal segmentation, BiLSTM with multi-head attention, and hierarchical MIL-based scoring to capture diverse temporal dependencies and improve anomaly localization in weakly labeled video data. Yang et al.31 proposed an attention-driven dual-branch framework that synchronized pseudo-label generation and self-training using self- and cross-attention-based snippet separation, denoised pseudo-labels, and multi-scale temporal feature interaction to enhance video anomaly detection performance. Sun et al.32 proposed an instance-modeling-focused framework that addressed anomaly contamination and dilution in MIL-based video anomaly detection by introducing dynamic segment merging to adaptively preserve abnormal segments and retrieval-augmented anomaly restoration to recover fine-grained anomalous features from a memory bank for improved detection performance. Zhang et al.33 proposed a dual-distillation-fusion framework that jointly leveraged contrastive distillation for filtering anomalous features and inpainting distillation for reconstructing normal patterns, with iterative self-training and joint optimization to enhance feature discrimination and anomaly detection performance in weakly supervised videos. Recent research has also leveraged vision and language alignment to enhance anomaly understanding. Wu et al.34 proposed VadCLIP, a vision-language framework for weakly supervised video anomaly detection that leverages the frozen CLIP model without fine-tuning. It employs a dual-branch architecture combining visual classification and video-text alignment to capture semantic relations. The framework further incorporates a local-global temporal adapter and prompt mechanisms to enhance temporal modeling and effective knowledge transfer. Wu et al.5 proposed STPrompt, a weakly supervised video anomaly detection and localization framework that learns spatio-temporal prompt embeddings using pre-trained vision-language models. The approach decomposes the task into temporal anomaly detection and spatial anomaly localization to reduce modeling complexity. It employs a spatial attention aggregation mechanism and a temporal adapter to capture motion priors, while a training-free LLM-based text prompting module enhances fine-grained spatial anomaly localization through concept-level patch retrieval. Yang et al.35 proposed TPWNG, a text-prompt-based weakly supervised video anomaly detection framework that leverages the multimodal capabilities of CLIP for pseudo-label generation and self-training. The method fine-tunes CLIP using ranking and distribution inconsistency losses to achieve domain adaptation while employing learnable text and normality visual prompts to enhance text–video alignment. Additionally, a temporal context self-adaptive learning module captures variable-length temporal dependencies, enabling more accurate and flexible modeling of anomalous events. Su et al.36 proposed VPE-WSVAD, a visual prompt exemplar-based framework for weakly supervised video anomaly detection designed to address intra-bag similarity and snippet entanglement issues. The method introduces a prompt retrieval network that extracts frame-level content aligned with scenario-aware visual exemplars, a proposal filtering mechanism that selects the most confident abnormal regions, and a prompt likelihood learning module that enhances the distinction between normal and abnormal representations. This architecture provides fine-grained reasoning of abnormal events in terms of when, where, and what, offering improved interpretability and adaptability to varying surveillance contexts. Qiu et al.37 proposed LOP-VAD, a vision-language-based framework that jointly leveraged learnable opposite yes and no prompts, temporal context attention, and prompt enhancement to model bidirectional semantics and temporal dependencies for improved anomaly localization in weakly labeled videos. Dev et al.38 proposed ReFLIP-VAD, a vision-language-based weakly supervised video anomaly detection framework that leveraged reparameterized learnable prompts, a dual-block architecture for classification and video-text alignment, and a Glimpse-Emphasize network to capture global and local temporal dependencies, enabling both coarse-grained and fine-grained anomaly detection with improved visual-language alignment. Liu et al.39 proposed VadCLIP++, a dynamic vision-language framework that leveraged learnable dynamic and static text prompts along with a spatial feature selection module to model frame-level temporal differences and spatial context, enabling enhanced spatiotemporal representation and improved anomaly detection in weakly labeled videos. Karim et al.40 propose SemVAD, a weakly supervised video anomaly detection framework that fuses semantic captions generated by VideoLLaMA 3 with CLIP visual features through a two-phase attention and convolution-based fusion module, augmented by a class-specific semantic alignment loss. Jia et al.41 introduce TSAVD-CLIP, which integrates temporal, spatial, and joint adapters directly into the CLIP image encoder to enable simultaneous spatiotemporal feature learning at the encoder level, coupled with a subspace alignment loss to maintain compatibility with the original CLIP text space. Shen et al.42 present TrCLIP-VAD, which enhances CLIP-based weakly supervised anomaly detection by rewriting video captions via LLaMA’s in-context learning to diversify textual inputs, and introduces an LGM-Mamba module combining local, global, and multi-scale temporal modeling for comprehensive spatiotemporal feature learning. Lim et al.43 propose ViCap-AD, a two-stage weakly supervised anomaly detection framework that generates video captions using CLIP4Clip and fuses the resulting BERT-encoded textual embeddings with visual features within a multiple instance learning framework augmented by a combined MIL ranking and binary cross-entropy loss. Motivation Although recent weakly supervised video anomaly detection methods have achieved strong performance through Multiple Instance Learning (MIL) and vision-language alignment, reliable anomaly localization under weak supervision remains challenging. In MIL-based learning, anomaly prediction depends on selecting a small set of highly confident temporal segments from long videos. However, when anomalous events exhibit only subtle deviations from normal patterns, inaccurate similarity estimation can lead to unstable Top-K instance selection and ambiguous anomaly ranking. At the same time, cross-modal alignment between temporal video segments and semantic descriptions must be learned using only video-level supervision, where the absence of frame-level annotations introduces additional uncertainty in learning reliable visual-text correspondence. Existing vision-language frameworks address this problem by aligning visual and textual representations in a shared embedding space, but they predominantly rely on cosine similarity in Euclidean space, which applies uniform distance scaling and may not effectively preserve fine-grained semantic variations between normal and anomalous events. Furthermore, standard WSVAD datasets provide only flat video-level labels and do not define any explicit hierarchical structure among event categories. Consequently, semantic relations must be inferred directly from cross-modal interactions rather than predefined structural assumptions. Motivated by recent advances in hyperbolic representation learning, we model visual and textual embeddings within the Poincaré ball manifold and compute similarity using negative Poincaré distance. The non-linear distance scaling induced by hyperbolic geometry improves separation between normal and anomalous instances and provides more consistent similarity estimation, leading to more stable MIL-based ranking and improved cross-modal alignment under weak supervision. Based on this formulation, the proposed PoinCLIP-VAD framework jointly performs anomaly classification and visual-text alignment within a unified hyperbolic space, enabling geometry-consistent cross-modal representation learning for weakly supervised video anomaly detection. Proposed framework As illustrated in Figure 1, the proposed PoinCLIP-VAD framework processes each input video by dividing it into short temporal segments and extracting frame-level visual features using a frozen CLIP image encoder augmented with a lightweight temporal adapter. This design preserves semantic priors from vision-language pretraining while capturing both short-term temporal dynamics and long-range contextual dependencies. The temporal adapter integrates a local transformer for fine-grained temporal modeling and a global graph convolutional module for capturing temporal coherence across frames. The resulting visual features are projected from Euclidean space into a hyperbolic manifold via an exponential mapping with a learnable curvature parameter based on the Poincaré ball formulation. In parallel, textual embeddings are constructed by combining learnable prompts with class tokens and are mapped into the same hyperbolic space to maintain consistency across modalities. Similarity between visual and textual representations is computed using negative Poincaré distance. The framework adopts a dual-branch architecture consisting of a Classification Block, which estimates frame-level anomaly scores using a Top-K Multiple Instance Learning strategy, and a Video-Text Alignment Block, which establishes correspondence between visual features and textual representations through hyperbolic distance-based alignment. Together, these components form a unified architecture for cross-modal representation learning and anomaly detection. Proposed framework. Problem statement Given a training set of videos , each video V is associated with a video-level label , where C denotes the total number of event categories encompassing both normal and anomalous behaviors. Each video is divided into N temporal segments, and high-level frame representations are extracted through a visual encoder, producing feature embeddings , where d is the feature dimension. The objective of weakly supervised video anomaly detection is to train a detector that, using only video-level supervision, learns to localize abnormal segments and infer their corresponding categories during testing. Encoders and temporal adapter Visual encoder The visual encoder is based on the pretrained CLIP image backbone, which jointly learns visual and linguistic representations through large-scale contrastive training on image–text pairs. Compared to traditional video feature extractors such as I3D, SlowFast, or ResNet-based models, CLIP offers a semantically grounded embedding space that bridges appearance features with high-level textual concepts, allowing effective transfer to unseen scenes and categories under weak supervision. This makes it particularly suitable for anomaly detection, where explicit frame-level supervision is unavailable and semantic generalization is essential. However, the native CLIP architecture operates on individual frames without modeling temporal continuity, causing its representations to overlook motion cues and contextual transitions crucial for distinguishing subtle or evolving anomalies. To address CLIP’s limitation in modeling temporal information, we integrate a lightweight temporal module, such as a Graph Convolution Network (GCN)44 or a Temporal Transformer34, to capture temporal dependencies effectively. This integration allows for efficient adaptation of CLIP’s pre-trained visual-textual knowledge to the Weakly Supervised Video Anomaly Detection (WSVAD) task. Text encoder The CLIP text encoder is designed to map language descriptions into a shared semantic space consistent with the visual embeddings. Each anomaly category is expressed as a short textual prompt that passes through the frozen CLIP text transformer, producing language-based embeddings that encapsulate both contextual meaning and relational similarity among classes. This joint representation allows intuitive alignment between visual evidence and linguistic concepts, even in weakly supervised settings. Visual temporal adapter Following VadCLIP34, we adopt a Visual Temporal Adapter as shown in Fig. 2 to bridge the gap between frame-wise CLIP representations and video-level temporal modeling. However, CLIP lacks explicit temporal modeling capability, limiting its ability to capture temporal dependencies across video frames. Specifically, the adapter consists of a local Transformer module and a global graph convolutional module operating sequentially on frame-level features extracted from the frozen CLIP image encoder. The local module captures short-range temporal dependencies by applying a window-based self-attention mechanism, where attention is restricted within overlapping temporal windows to preserve fine-grained temporal continuity while maintaining computational efficiency. Building upon this, the global module employs a lightweight GCN to model long-range dependencies by constructing adjacency relationships based on feature similarity and temporal positional distance, enabling effective information propagation across distant frames. Residual connections are incorporated in both modules to prevent over-smoothing and preserve discriminative feature representations. This design allows the model to jointly encode short-term dynamics and long-range contextual structure, producing temporally enriched visual features that are crucial for accurate anomaly localization and robust cross-modal alignment under weak supervision. Illustration of Visual Temporal Adapter. Prompt-integrated dual block framework Dual-block architecture We adopt a dual-block architecture, comprising a classification block and a visual-text alignment block. After Temporal Modelling, the video feature is fed to a fully connected layer to obtain the final video feature which is then scaled by a learnable factor and projected into the Poincaré ball manifold using the exponential map at the origin with curvature c. This transformation maps Euclidean embeddings into a hyperbolic space, introducing non-linear distance scaling that improves separability between normal and anomalous instances under weak supervision, where subtle differences are difficult to distinguish in Euclidean space. The mapping is denoted as where denotes the exponential mapping and represents the hyperbolic embeddings on the manifold. In the Classification block, we feed into the hyperbolic classifier to obtain the Anomaly Confidence . Specifically, the anomaly score for the i-th frame is given by where denotes the Poincaré distance, is a learnable prototype mapped onto the manifold via exponential mapping, b is a bias term, and is the sigmoid activation. The distance is defined as where denotes Möbius addition on the Poincaré ball Manifold. The use of geodesic distance establishes the decision boundary based on manifold-aware separation, improving robustness to subtle feature variations. Although the classifier formulation is inspired by HyperVD45, its role in the proposed framework is fundamentally different. Rather than functioning only as a standalone anomaly prediction module, the predicted anomaly scores are further utilized for multimodal prompt aggregation and visual-text alignment within the dual-block framework. This establishes a direct interaction between coarse-grained anomaly classification and fine-grained cross-modal representation learning, allowing the classification branch to guide semantic alignment between visual features and textual representations under weak supervision. In the visual-text alignment (VTA) block, textual labels such as abuse, riot, and fighting are encoded into class embeddings using a pre-trained CLIP text encoder, which remains frozen to preserve its rich linguistic priors. To ensure geometric consistency between modalities, these textual embeddings are projected onto the Poincaré ball manifold via the exponential mapping. Once both frame-level visual features and class embeddings reside in the same hyperbolic space, similarity is computed using the negative Poincaré distance. This results in an alignment map , where m is the number of text labels. In the Alignment Block, each input text label represents a class of abnormal events, thus naturally achieving fine-grained WSVAD. Context-adaptive prompt In weakly supervised video anomaly detection, textual labels are words or phrases which are too succinct to summarize abnormal events very well. To enrich these limited textual cues and improve the adaptability of text embeddings across different anomaly contexts, we design a context-adaptive prompt inspired by the prompting paradigm in VadCLIP34. Each discrete label is first tokenized through the CLIP tokenizer to obtain the initial class token where Label is the discrete text label, eg.; “fighting”, “shooting”, “road accident”, etc. We then concatenate the initial text embedding with the learnable prompt that contains l context tokens to construct a complete sentence token, which serves as the input to the text encoder. Here, we place the class token at the middle of the sequence. Then, this token sequence is added to the positional embedding to obtain positional information, and finally, the text encoder of CLIP takes as input and generates the class embedding . To ensure geometric consistency with the hyperbolic visual space, the embedding is scaled by a learnable parameter and projected onto the Poincaré ball manifold through the exponential map as follows: where denotes the class embedding residing on the Poincaré ball manifold. Multimodal prompt To further strengthen the representational capacity of textual labels for abnormal event categories, we incorporate visual context to refine the class embedding, since visual contexts can make the succinct text labels more accurate. To this end, we propose a multimodal prompt that focuses on the visual embeddings in abnormal segments and aggregates these embeddings as the video-level prompt for class embeddings. We first use the anomaly confidence A, obtained from the Classification Block, as the anomaly attention, then compute the video-level prompt as follows: where denotes the tangent-space normalization, and is the anomaly focus visual prompt. We then add to the class embedding and obtain the final instance- specific class embedding by a feed-forward network layer and a skip connection, which can be formulated as This implementation allows the hyperbolic class embeddings to extract the most relevant visual context from videos while preserving curvature consistency. Finally, with and both defined on the Poincaré ball manifold, we calculate the match similarities between all class embeddings and frame-level visual features to obtain the alignment map using the negative Poincaré distance: where denotes Möbius addition on the Poincaré ball manifold and is the temperature parameter controlling similarity sharpness. Objective function Classification block (CLS block) In the classification block, we employ the widely adopted Top-K selection mechanism to select the K highest anomaly confidences in both normal and abnormal videos as the video-level predictions. Operating in hyperbolic space introduces non-linear distance scaling, which reduces ambiguity between closely related instances and improves the stability of the ranking process under weak supervision. Then we use the binary cross-entropy between video-level predictions and ground-truth to compute classification loss . Visual-text alignment block (VTA block) The Visual-Text Alignment Block establishes correspondence between visual features and class embeddings within a unified hyperbolic space. For each video sequence, an alignment map is constructed to represent the similarity between frame-level video features and all class embeddings using the negative Poincaré distance. For each row, we select the Top-K similarities and compute the average to measure the alignment degree between this video and the current class. Then we obtain a vector , that represents the similarity between this video and all classes, with the video and its paired textual label attaining the highest similarity score among others. To achieve this, the multi-class prediction is first computed as follows. where is the prediction with respect to the class and refers to the temperature hyper-parameter for scaling. Finally, the alignment loss is computed by the cross-entropy. Since standard WSVAD datasets provide only flat category annotations, no predefined semantic hierarchy is assumed, and relational structure is instead learned implicitly from the embedding geometry. To further model semantic relations within the hyperbolic manifold, localized entailment regions are constructed in the Poincaré space following the entailment cone formulation introduced in PoinCLIP. Unlike conventional Euclidean alignment, the entailment objective regularizes the relative organization of embeddings by constraining semantically related visual and textual representations to occupy geometrically consistent regions within the hyperbolic space. Rather than enforcing a fixed category-level hierarchy, the entailment mechanism captures asymmetric semantic relations through geometry-aware inclusion constraints. The entailment loss between a video embedding and a text embedding is formulated as where is the extent of the angle between embeddings, denotes the aperture of the entailment cone, and is a regularization coefficient that constrains excessive cone expansion. To further enhance separation between normal and abnormal classes, a contrastive separation objective is introduced to maximize the geodesic margin between the normal embedding and abnormal embeddings : where denotes the Poincaré distance and specifies the margin threshold. This objective complements the entailment regularization by improving inter-class separation while preserving the geometry-aware semantic organization of the embedding manifold. Finally, the complete learning objective combines classification, alignment, entailment, and contrastive components into a unified optimization function: where and control the relative contribution of entailment and contrastive regularization. Experiments and results Experimental datasets and evaluation metrics Experimental datasets We evaluate the proposed framework on two publicly available benchmark datasets, UCF-Crime8 and XD-Violence46. The UCF-Crime dataset is a large-scale compilation of real-world surveillance footage tailored for weakly supervised video anomaly detection. It contains approximately 128 hours of video across 1,900 clips representing 13 different categories of anomalous activities. Among these, 1,610 videos with video-level annotations are used for training, while the remaining 290 clips, which include frame-level ground truth, are reserved for evaluation. The XD-Violence dataset serves as an extensive benchmark for studying violent event detection. It encompasses around 217 hours of video collected from diverse sources such as films, online media, surveillance recordings, and real-world scenarios. The dataset includes a total of 4,754 clips divided into six categories of violent or anomalous behavior. Out of these, 3,954 videos with video-level annotations are employed for training, and 800 videos containing frame-level labels are used for testing. Evaluation metrics For the coarse-grained weakly supervised video anomaly detection task, we follow previous studies and evaluate performance using the frame-level Average Precision (AP) metric on the XD-Violence dataset and the frame-level Area Under the Curve (AUC) on UCF-Crime. For the fine-grained WSVAD setting, we adopt the standard temporal action detection protocol and compute the mean Average Precision (mAP) at several intersection-over-union (IoU) thresholds. Specifically, IoU values from 0.1 to 0.5 are used with an interval of 0.1, and their mean is reported as the final AVG score. It should be noted that mAP is calculated only over the abnormal videos included in the test set. Implementation details The proposed PoinCLIP-VAD framework is implemented in PyTorch and trained on a single NVIDIA RTX 4090 GPU using mixed precision training. The visual sampling rate is set to 24 fps for UCF-Crime and 30 fps for XD-Violence, while visual features are extracted using a sliding window of 16 frames. The pretrained CLIP ViT-B/16 is adopted as the visual encoder, and the CLIP Transformer is used as the text encoder, both initialized from large-scale image-text pretraining. The backbone encoders remain frozen during training, while only the temporal adapter, prompt parameters, and hyperbolic projection layers are optimized. Each video is divided into non-overlapping 16-frame clips with a stride of 8, and the extracted frame-level features are processed through the proposed Visual Temporal Adapter to capture both local and global temporal dependencies. The resulting visual and textual embeddings are projected into a shared Poincaré ball manifold through exponential mapping at the origin, where the learnable curvature parameter is constrained within to ensure stable hyperbolic optimization. The modality scaling factors are initialized as , and cross-modal similarity is computed using negative Poincaré distance. The overall optimization objective combines binary cross-entropy loss, hyperbolic alignment loss, entailment regularization loss, and curvature-regularized contrastive loss. The model is optimized using Adam Optimizer with a batch size of 64, weight decay of , and an initial learning rate of . A cosine annealing scheduler is adopted during training, and the temperature parameter is fixed at 0.07. For the loss weighting parameters, is set to 0.1 for UCF-Crime and for XD-Violence, while the curvature regularization coefficient is fixed at for both datasets. The proposed framework demonstrates stable convergence within 30 epochs under the hyperbolic optimization setting. For the MIL-based learning strategy, the value of in Top-K selection is defined as for positive bags and for negative bags, where denotes the temporal length of the input features. Experimental results The effectiveness and generalization of the proposed PoinCLIP-VAD framework are validated through comprehensive experiments conducted on the UCF-Crime8 and XD-Violence46 datasets, covering both coarse-grained and fine-grained anomaly detection settings. Performance comparison on UCF Crime For the UCF-Crime dataset, the coarse-grained results summarized in Table 1 show that PoinCLIP-VAD achieves an AUC of 90.62%, outperforming all existing Euclidean-space-based approaches. Compared to VadCLIP (86.70%), the proposed framework achieves a relative improvement of 3.92%, demonstrating the effectiveness of hyperbolic representation learning for weakly supervised anomaly detection. Furthermore, PoinCLIP-VAD consistently surpasses several recent state-of-the-art CLIP-related methods. Although these methods improve semantic alignment through prompt learning, temporal adaptation, and enhanced vision-language interaction, their Euclidean formulations still exhibit limitations in maintaining consistent similarity estimation and discriminative separation between normal and anomalous instances. Consequently, they may suffer from instability in MIL-based Top-K selection and weaker anomaly discrimination across temporally similar events. In contrast, the curvature-aware architecture of PoinCLIP-VAD enables more reliable geometry-consistent similarity estimation and stronger cross-modal alignment, resulting in improved stability of MIL-based ranking and enhanced anomaly discrimination. The consistent improvement over both conventional WSVAD approaches and recent CLIP-based variants further validates the effectiveness of integrating hyperbolic geometry into weakly supervised vision-language anomaly detection. Table 1Comparison of coarse-grained anomaly detection results on the UCF-Crime dataset.SupervisionMethodFeature SpaceAUC (%)Un-supSVM BaselineEuclidean50.10Hasan et al.47Euclidean51.20OCSVM48Euclidean63.20Wk-SupIBL49Euclidean78.66Motion-Aware9Euclidean78.66Deep-MIL8Euclidean75.42HL-Net50Euclidean82.13GCN44Euclidean82.12MS-BSAD51Euclidean83.54Ju et al.52Euclidean84.22LA-Net53Euclidean84.87CRFD54Euclidean84.89RTFM4Euclidean85.28NG-MIL55Euclidean85.63AnomalyCLIP56Euclidean86.50DMU57Euclidean86.75UB-MIL58Euclidean86.97VadCLIP34Euclidean86.70ViCap-AD43Euclidean87.20TSAVD-CLIP41Euclidean87.21SemVAD40Euclidean88.48CLIP-TSA59Euclidean87.58ST-PROMPT5Euclidean88.08VadCLIP++39Euclidean88.12ReFLIP-VAD38Euclidean88.58TrCLIP-VAD42Euclidean88.59Wk-SupPoinCLIP-VAD (Ours)Hyperbolic90.62 Comparison of coarse-grained anomaly detection results on the UCF-Crime dataset. To further examine temporal precision, we analyze fine-grained anomaly detection performance, as reported in Table 2. PoinCLIP-VAD consistently outperforms all competing methods at every IoU threshold, achieving a mean Average Precision (mAP) of 10.76%, which is 4.08% higher than VadCLIP (6.68%). These results suggest that the hyperbolic feature space provides more consistent similarity across temporal segments, leading to improved localization of subtle anomalies under weak supervision. Overall, PoinCLIP-VAD achieves state-of-the-art results on UCF-Crime across both coarse and fine evaluations, demonstrating that geometry-consistent similarity modeling improves ranking stability and alignment quality within the embedding space. Table 2Comparison of fine-grained anomaly detection results on the UCF-Crime dataset.Method0.10.20.30.40.5AVGRandom0.210.140.040.020.010.08Sultani et al.85.734.412.691.931.443.24AVVD5010.277.016.253.423.296.05VadCLIP3411.727.836.404.532.936.68ReFLIP-VAD3814.2310.349.327.546.819.62PoinCLIP-VAD (Ours)15.4811.0210.348.278.7110.76 Comparison of fine-grained anomaly detection results on the UCF-Crime dataset. Performance comparison on XD Violence Our proposed PoinCLIP-VAD framework demonstrates strong performance on the XD-Violence dataset, as summarized in Table 3. Under the weakly supervised setting, PoinCLIP-VAD achieves an Average Precision (AP) of 86.93%, outperforming all existing Euclidean-space-based approaches. Compared with VadCLIP (84.51%), the proposed framework achieves an improvement of 2.42%, highlighting the effectiveness of hyperbolic representation learning for complex violent-event analysis. Furthermore, PoinCLIP-VAD consistently surpasses several recent CLIP-based frameworks, including SemVAD, ST-PROMPT, ViCap-AD, VadCLIP++, ReFLIP-VAD, and TrCLIP-VAD. Although these methods improve multimodal understanding through prompt optimization and temporal adaptation, violent-event scenarios often involve abrupt motion variations and visually ambiguous transitions between normal and abnormal activities, making semantic separation challenging in Euclidean space. In contrast, the hyperbolic embedding space of PoinCLIP-VAD provides enhanced representational separation for complex temporal patterns, enabling clearer discrimination of subtle anomaly transitions across video segments. The curvature-aware formulation further improves cross-modal alignment between visual representations and anomaly-aware textual prompts, resulting in more robust anomaly localization and stable MIL-based ranking. These consistent improvements validate the effectiveness of integrating hyperbolic geometry into weakly supervised vision-language anomaly detection. Table 3Comparison of coarse-grained anomaly detection results on the XD Violence dataset.SupervisionMethodFeature SpaceAP (%)Un-supOCSVM48Euclidean28.63Hasan et al.47Euclidean31.25BaselineEuclidean50.80Wk-SupRTFM4Euclidean78.27HL-Net50Euclidean78.10MS-BSAD51Euclidean78.92Deep-MIL8Euclidean75.18CRFD54Euclidean75.90NG-MIL55Euclidean78.51UB-MIL58Euclidean81.20LA-Net53Euclidean80.72Ju et al.52Euclidean76.57AnomalyCLIP56Euclidean78.55CLIP-TSA59Euclidean82.17DMU57Euclidean82.41Que et al.60Euclidean83.82VadCLIP34Euclidean84.51SemVAD40Euclidean84.66ST-PROMPT5Euclidean84.75ViCap-AD43Euclidean85.02VadCLIP++39Euclidean85.03ReFLIP-VAD38Euclidean85.81TrCLIP-VAD42Euclidean86.05Wk-SupPoinCLIP-VAD (Ours)Hyperbolic86.93 Comparison of coarse-grained anomaly detection results on the XD Violence dataset. For the fine-grained WSVAD on the XD-Violence dataset, we compare the proposed PoinCLIP-VAD with recent vision-language baselines, including VadCLIP and other representative methods. As shown in Table 4, this task is notably more challenging than the coarse-grained setting, as it requires accurate multi-category classification along with precise temporal boundary localization. Despite the increased difficulty, PoinCLIP-VAD consistently outperforms all competing approaches across different IoU thresholds. Specifically, it achieves an average mAP of 28.11%, yielding an absolute improvement of 3.41% over VadCLIP. The performance gain becomes more evident at higher IoU levels, indicating that more consistent similarity across temporal segments improves the localization of subtle anomalies under weak supervision. Overall, these results demonstrate that geometry-consistent similarity modeling improves ranking stability and cross-modal alignment, leading to superior fine-grained anomaly localization on XD-Violence. Table 4Comparison of fine-grained anomaly detection results on the XD Violence dataset.Method0.10.20.30.40.5AVGBaseline1.820.920.480.230.090.71Sultani et al.822.7215.579.986.203.7811.65AVVD5030.5125.7520.1814.839.7920.21VadCLIP3437.0330.8423.3817.9014.3124.70ReFLIP-VAD3839.2433.4527.7120.8617.2227.36PoinCLIP-VAD (Ours)39.6834.4228.0721.1319.2428.11 Comparison of fine-grained anomaly detection results on the XD Violence dataset. Qualitative analysis Figure 3 presents the qualitative visualization results on the UCF-Crime and XD-Violence datasets, respectively. The blue solid lines indicate the frame-level anomaly scores predicted by PoinCLIP-VAD, while the light orange shaded areas correspond to the ground-truth abnormal temporal segments. It can be observed that the proposed model accurately localizes the onset and duration of abnormal events such as explosions, shootings, riots, abuse, and fighting, with clear score peaks aligning with the annotated regions. The smooth transitions in normal segments further demonstrate the model’s ability to suppress false positives and maintain temporal stability. However, certain failure cases are also observed, such as shoplifting, where the abnormal behavior unfolds gradually with minimal motion and weak visual saliency, resulting in low anomaly confidence. Additionally, false positives arise in contextually ambiguous scenarios, such as aggressive interactions during sports activities (e.g., ice hockey collisions), where visually similar patterns resemble violent events despite being normal within the given context. These limitations stem from the reliance on strong semantic cues under weak supervision and the lack of explicit contextual reasoning, highlighting the challenges in capturing fine-grained temporal dynamics and distinguishing semantically similar yet contextually different activities. Frame-level scores and violence localization examples from the XD-Violence and UCF-Crime datasets. We visualize the learned embedding distributions on the XD-Violence dataset in Figure 4. The left panel presents the Euclidean t-SNE projection of the original CLIP embeddings, while the right panel illustrates the Poincaré disk representation of the hyperbolic embeddings learned by PoinCLIP-VAD. Although the Euclidean visualization primarily reflects cluster separability in Euclidean space, the hyperbolic representation further reveals geometry-aware semantic organization within the embedding manifold. In particular, generic and ambiguous samples remain concentrated near the center of the Poincaré disk, whereas more discriminative anomaly instances progressively expand toward the boundary. Additionally, the dashed geodesic trajectories illustrate asymmetric cross-modal alignment between text anchors and visual embeddings, consistent with the entailment formulation defined in Eq.11. Compared with the original CLIP features, the proposed hyperbolic embeddings exhibit improved local consistency, reduced semantic overlap among challenging anomaly categories, and more structured manifold organization, demonstrating that the proposed curvature-aware optimization captures meaningful semantic relations beyond conventional Euclidean cluster separation under weak supervision. Visualization of learned embeddings on the XD-Violence dataset. (a) t-SNE projection of Euclidean CLIP embeddings shows cluster separability in Euclidean space. (b) Poincaré disk visualization of hyperbolic embeddings learned by PoinCLIP-VAD. Embeddings organize radially in the hyperbolic manifold, where generic and ambiguous samples concentrate near the center, while more discriminative anomaly instances spread toward the boundary. Ablation studies Effectiveness of Dual Block As presented in Table 5, the proposed framework demonstrates the progressive contribution of each component on the XD-Violence dataset. When only the CLS-Block is used, the model follows a coarse-grained classification paradigm and already achieves a competitive AP of 81.07%, confirming the strong discriminative capacity of the hyperbolic classifier for anomaly identification. In contrast, using only the VTA-Block, which performs fine-grained video-text alignment under weak supervision, results in lower accuracy (69.82%) due to the absence of coarse contextual cues. When both blocks are optimized jointly, the overall performance improves to an AP of 76.48%, indicating that the interaction between classification and alignment learning produces more robust and stable feature representations. Incorporating the CA-Prompt further enhances language adaptability and semantic consistency, leading to an AP of 82.09%, while the addition of the multimodal prompt introduces cross-modal contextual reasoning and elevates the performance to 84.11%. Finally, the complete configuration that integrates all components achieves an AP of 86.93%, surpassing existing approaches and validating that the complementary synergy between coarse-grained classification and fine-grained alignment effectively enhances anomaly representation and detection performance in complex video scenarios. Table 5Effectiveness of dual block.CLS-BlockVTA-BlockCA-PromptM-PromptAP (%)✓81.07✓69.82✓✓76.48✓✓✓82.09✓✓✓84.11✓✓✓✓86.93 Effectiveness of dual block. Computational analysis We further evaluate the computational efficiency and deployment characteristics of PoinCLIP-VAD, as summarized in Table 6. For a fair and consistent comparison, all methods are evaluated using their official implementations under the same experimental setup on an NVIDIA RTX 4090 GPU, and beyond commonly reported proxy indicators such as parameter count and FLOPs, we include practical runtime metrics, including inference latency, throughput, and GPU memory consumption, to provide a realistic assessment of deployment feasibility. PoinCLIP-VAD achieves superior detection performance, attaining 90.62% AUC on UCF-Crime and 86.93% AP on XD-Violence, while maintaining a compact architecture with 25M parameters and the lowest computational cost of 11.6G FLOPs among all compared methods, suggesting that the performance gains are associated with improved representation efficiency rather than increased model complexity. In terms of runtime efficiency, the model achieves a latency of 7 ms and a throughput of 102 clips per second, maintaining competitive runtime efficiency relative to existing CLIP-based approaches while consistently delivering higher accuracy. Notably, the hyperbolic projection and distance computation operate in a low-dimensional embedding space, introducing only limited additional computational overhead while improving representation separability between normal and anomalous instances and contributing to more stable MIL-based ranking under weak supervision. Overall, these results indicate that PoinCLIP-VAD effectively balances model complexity, computational efficiency, and detection performance, highlighting that curvature-aware geometry provides task-level benefits for anomaly discrimination without assuming any predefined hierarchical structure in the data. Table 6Comparison of computational complexity, inference efficiency, and performance across different methods.MethodPerformanceInferenceComplexityUCF-Crime AUC (%)XD-Violence AP (%)Latency (ms) Throughput (clips/s) Params (M)FLOPs (G)Memory (MB)SSRL6183.2077.3535281912153200RTFM484.3077.812835281872400MIST-I3D6283.6779.18166231462100MIST-C3D6283.1279.58146885392300STPrompt588.0882.41128532451800VadCLIP3488.0284.5191124013.81500PoinCLIP-VAD (Ours)90.6286.9371022511.61400 Comparison of computational complexity, inference efficiency, and performance across different methods. Effect of similarity metric and geometric consistency To isolate the contribution of hyperbolic geometry, we replace the Poincaré distance with cosine similarity while keeping all other components of the framework unchanged. This modification enables a controlled comparison to determine whether the observed performance gains arise from the underlying geometry rather than architectural factors. The results are summarized in Table 7. The cosine-based variant achieves an AUC of 85.14% on the UCF-Crime dataset and an AP of 81.37% on the XD-Violence dataset, compared to an AUC of 90.62% and an AP of 86.93% obtained using the Poincaré distance. Table 7Effect of similarity metric on UCF-Crime and XD-Violence datasets.Similarity metricUCF-crime (AUC %)XD-violence (AP %)Cosine similarity85.1481.37Poincaré Distance90.6286.93 Effect of similarity metric on UCF-Crime and XD-Violence datasets. Beyond numerical performance, the improvement primarily originates from geometry-consistent similarity modeling rather than architectural modifications or predefined hierarchical assumptions. Hyperbolic space, a Riemannian manifold with constant negative curvature, exhibits non-linear scaling in the Poincaré ball, providing exponentially increasing representational capacity toward the boundary and enabling better separation between semantically similar instances. Accordingly, PoinCLIP employs hyperbolic distance within the contrastive objective so that similarity computation remains naturally aligned with the underlying embedding geometry. In contrast, cosine similarity assumes a flat Euclidean space and ignores the manifold structure, introducing a geometric mismatch that produces less reliable similarity estimation. This inconsistency can lead to ambiguous temporal ranking and unstable Top-K selection during weakly supervised learning, particularly when subtle anomalies must be distinguished under limited supervision. By preserving and amplifying fine-grained inter-instance differences, hyperbolic distance enables more stable anomaly discrimination and stronger cross-modal alignment between visual and textual representations. Therefore, the observed performance gains arise from geometry-aware similarity estimation and improved representation separability rather than explicit assumptions about category-level hierarchies. Curvature convergence analysis To investigate the geometric behavior of the proposed framework, we analyze the evolution of the learnable curvature parameter during training. The parameter , which controls the negative geometry of the Poincaré manifold, is learned dynamically and constrained within the range to ensure numerical stability, where the corresponding sectional curvature of the hyperbolic space is represented as . The progression of curvature values across training epochs for both datasets is presented in Table 8. It can be observed that the curvature increases consistently during optimization and stabilizes after approximately 25–30 epochs, indicating stable and reliable convergence behavior. Table 8Curvature convergence behavior across training.EpochUCF-Crime cSectional Curvature XD-Violence cSectional Curvature 10.18− 0.180.12− 0.1250.34− 0.340.19− 0.19100.51− 0.510.25− 0.25200.66− 0.660.29− 0.29250.70− 0.700.30− 0.3030 (Converged)0.72− 0.720.31− 0.31 Curvature convergence behavior across training. A clear dataset-dependent trend is observed in the learned curvature dynamics. For UCF-Crime, the curvature increases from 0.18 to 0.72 , resulting in a more negatively curved space that increases separation between normal and anomalous instances and contributes to more stable MIL-based ranking, consistent with the relatively coarse and separable anomaly patterns in the dataset. In contrast, XD-Violence converges to a lower curvature of 0.31 , producing a geometry closer to Euclidean space that retains finer local relationships and is better suited to the diverse and overlapping anomalies present in the dataset. These observations suggest that the learned curvature adapts according to dataset complexity rather than remaining fixed, where higher curvature improves instance-level discrimination while lower curvature supports stable representation learning for more complex anomaly distributions. Overall, the convergence behavior demonstrates the effectiveness of dynamic curvature learning in adapting the embedding geometry to different data characteristics. Effectiveness of different hyper-parameter settings As reported in Table 9, the best performance is obtained when the temperature parameter () is set to 0.07, achieving an AUC of 90.62% on UCF-Crime and an AP of 86.93% on XD-Violence. This setting consistently outperforms all other configurations, offering the most stable trade-off between inter-class discrimination and intra-class cohesion within the hyperbolic contrastive embedding space. Notably, using leads to gains of up to 1.71% in AUC and 2.17% in AP over the weakest configurations. These findings indicate that adopting a moderate temperature effectively controls the gradient scaling in the contrastive loss, enabling more compact yet well-separated feature distributions across both benchmarks. Table 9Performance comparison of different temperature coefficients on UCF-Crime and XD-Violence datasets.0.010.030.050.070.090.1UCF-Crime (AUC %)88.9189.8489.2790.6289.5689.31XD-Violence (AP %)84.7685.4385.2786.9386.0285.68 Performance comparison of different temperature coefficients on UCF-Crime and XD-Violence datasets. Table 10 presents the sensitivity analysis for the curvature-regularization coefficient , which balances the contrastive and hyperbolic alignment losses. As increases from to , both AUC and AP improve steadily, showing that moderate curvature regularization enhances manifold smoothness and discriminative separation. Beyond , over-regularization causes performance to decline due to excessive constraint on manifold curvature. Empirically, yields the best performance for UCF-Crime, while performs optimally for XD-Violence, aligning with the differing curvature sensitivities and dataset complexities. Table 10Effectiveness of the curvature-regularization coefficient on UCF-Crime and XD-Violence datasets.CurvatureUCF-CrimeXD-ViolenceRegularizationAUC(%)AP(%)Weak87.2484.12Moderate88.3385.17Optimal (XD-Violence)90.0286.93Strong89.7585.61Over-regularized88.5683.920.1Optimal (UCF-Crime)90.6286.14 Effectiveness of the curvature-regularization coefficient on UCF-Crime and XD-Violence datasets. To examine the effect of the entailment-cone weight () and the contrastive regularization coefficient () in Eq.13, we conduct an ablation study on both datasets while keeping other parameters fixed (, ). The value of is shared across datasets to maintain curvature consistency, whereas is tuned per dataset to adapt to varying label noise and task granularity. The experimental outcomes presented in Table 11 show that maintaining a unified across datasets ensures stable curvature-aware embedding, while dataset-specific tuning of effectively balances latent semantic inclusion and contrastive separation within the embedding space rather than relying on explicitly defined hierarchical relationships. For the large and noisy XD-Violence dataset, a smaller prevents over-separation and achieves the highest fine-grained performance (AUC: 90.02%, AP: 86.93%). In contrast, the binary UCF-Crime dataset benefits from a stronger , which enhances inter-class margins and yields optimal coarse-level anomaly discrimination (AUC: 90.62%, AP: 86.14%). These findings confirm that should remain geometry-invariant across datasets, whereas requires dataset-level adaptation for optimal generalization. Table 11Ablation study of and on UCF-Crime and XD-Violence datasets.DatasetAUC(%)AP(%)XD-Violence87.2484.12XD-Violence88.3385.17XD-Violence90.0286.93XD-Violence89.7585.610.1UCF-Crime90.6286.140.1UCF-Crime88.5683.92 Ablation study of and on UCF-Crime and XD-Violence datasets. Effectiveness of context length The experimental results presented in Table 12 demonstrate that the context length l, defined as the number of learnable context tokens, plays a decisive role in shaping the representational capacity and semantic richness of textual embeddings and their effectiveness in cross-modal alignment. When a shorter context length of is employed, the model exhibits comparatively lower performance, as the limited number of context tokens constrains the representational capacity of the prompt, thereby restricting its ability to capture fine-grained semantic variations between visual features and textual descriptions. Increasing the context length to leads to a substantial improvement, achieving an AUC of 90.62% on the UCF-Crime dataset and an AP of 86.93% on the XD-Violence dataset. This improvement can be attributed to enhanced contextual modeling, which enables the text encoder to generate more informative and discriminative embeddings, thereby strengthening the alignment between visual and textual modalities within the shared embedding space. Table 12Effect of context length on UCF-Crime and XD-Violence datasets.Context Length (l)UCF-Crime (AUC %)XD-Violence (AP %)889.7685.841290.6286.931690.3486.512089.9886.07 Effect of context length on UCF-Crime and XD-Violence datasets. When the context length is further increased to , the performance exhibits a slight decline, which can be attributed to the introduction of partially redundant context tokens that diffuse the attention distribution of the text encoder across less informative components, thereby weakening semantic focus. This effect becomes more pronounced at , where additional tokens contribute minimal semantic value while increasing representational redundancy, ultimately reducing the precision of feature alignment. This degradation is further amplified in the hyperbolic embedding space, where similarity computation is inherently sensitive to perturbations, and even minor variations in textual representations can result in disproportionately larger deviations in distance-based similarity estimation. Effectiveness of class-wise AUC and AP results The class-wise AUC results on the UCF-Crime dataset for the top three performing methods, UB-MIL58, CLIP-TSA59, and PoinCLIP-VAD (Ours), are shown in Fig. 5. It is evident that PoinCLIP-VAD consistently achieves the highest AUC values across nearly all anomaly categories, showing its strength in capturing the intrinsic hierarchy of anomalous events through a hyperbolic feature space. Significant improvements are observed in categories such as Explosion, Robbery, and Road Accident, where the curvature-aware classifier provides better separation of complex motion and contextual cues. Even in visually ambiguous categories such as Shoplifting and Stealing, where motion cues are subtle, and scene variations are minimal, PoinCLIP-VAD maintains a clear margin over CLIP-TSA and UB-MIL. These results validate that the dual branch design, which combines coarse-grained classification and curvature-aware alignment, improves feature discriminability while preserving structural consistency. Overall, PoinCLIP-VAD achieves an average AUC of 90.62%, surpassing CLIP-TSA (87.58%) and UB-MIL (86.97%), setting a new benchmark in coarse-grained anomaly detection. Class-wise AUC results of the top three methods on the UCF-Crime dataset. The class-wise average precision results on the XD-Violence dataset for CLIP-TSA59, DMU57, and PoinCLIP-VAD (ours) are shown in Fig. 6. PoinCLIP-VAD achieves the highest AP across all six violence categories, especially in Abuse, Explosion, and Riot, where temporal consistency and multimodal reasoning play a key role. In the Abuse category, PoinCLIP-VAD attains an AP of about 87%, outperforming CLIP-TSA (84%) and DMU (80%), while in Car Accident, it achieves around 88%, showing stronger temporal alignment and better multimodal context understanding. The model’s curvature-regularized optimization and Top-K MIL strategy effectively align visual and textual features, reducing ambiguity and enhancing fine-grained discrimination. These consistent improvements show that PoinCLIP-VAD enhances frame-level precision and temporal stability, achieving an overall AP of 86.93% and setting a new standard for fine-grained multimodal anomaly detection. Class-wise AUC results of the top three methods on the XD-Violence dataset. Impact of hyperbolic geometry and manifold selection The impact of the underlying hyperbolic manifold is analyzed through a controlled comparison between the Poincaré ball and Lorentz (hyperboloid) models, as summarized in Table 13. Both formulations achieve strong and consistent performance, indicating that the primary gains originate from hyperbolic geometry itself rather than the specific manifold parameterization. The Lorentz-based variant achieves an AUC of 90.21% on UCF-Crime and an AP of 86.58% on XD-Violence, while the Poincaré model attains slightly higher scores of 90.62% and 86.93%, respectively. Although both are isometric realizations of the same negatively curved space, they differ in geometric characteristics and optimization behavior. The Lorentz model represents hyperbolic space as a hyperboloid in Minkowski space and is known for numerical stability due to closed-form geodesics and the absence of boundary effects, whereas the Poincaré ball provides a conformal representation that preserves local angular relationships, which is advantageous for similarity-based learning. Table 13Effect of hyperbolic manifold choice on the experimental datasets.Hyperbolic ModelUCF-Crime (AUC %)XD-Violence (AP %)Lorentz90.2186.58Poincaré Ball90.6286.93 Effect of hyperbolic manifold choice on the experimental datasets. The marginal advantage of the Poincaré model lies in its stronger compatibility with the similarity-driven objectives of the proposed cross-modal learning framework. Its conformal structure preserves local neighborhood relationships and angular consistency, enabling more reliable cross-modal similarity estimation and improved discrimination between normal and anomalous instances. The negative Poincaré distance further provides geometry-consistent similarity modeling that better captures subtle semantic variations and improves the stability of similarity computation within the embedding space. Moreover, the bounded nature of the Poincaré ball leads to a more interpretable embedding organization, where semantically generic representations remain near the center while more discriminative anomaly patterns progressively expand toward the boundary, without requiring predefined hierarchies. Although the Lorentz model provides favorable numerical stability in hyperbolic optimization, it relies on a Minkowski inner product where angular relationships are not explicitly preserved, which can introduce slight inconsistencies in similarity estimation and cross-modal alignment. Prior studies12,63 have established that hyperbolic space offers exponentially greater representational capacity for modeling complex semantic structures. Although both manifolds are theoretically equivalent, the Poincaré formulation integrates more naturally with exponential mapping and contrastive similarity computation, resulting in more stable semantic alignment and enhanced anomaly discrimination. Effectiveness of loss functions As presented in Table 14, we systematically investigate the contribution of each loss component and observe consistent performance degradation upon their removal across both benchmarks, which can be attributed to the distinct functional role each term plays in the proposed framework. Using only the classification loss yields a baseline performance of 84.37% AUC on the UCF-Crime dataset and 79.92% AP on the XD-Violence dataset, reflecting the inherent limitation of coarse video-level supervision in capturing fine-grained semantic variations. Excluding the alignment loss leads to a drop to 86.12% AUC on the UCF-Crime dataset and 81.45% AP on the XD-Violence dataset, as the MIL-based video-text matching mechanism is no longer enforced, thereby weakening the semantic correspondence between visual features and anomaly categories. Similarly, removing the contrastive loss results in reduced performance of 89.05% AUC on the UCF-Crime dataset and 85.21% AP on the XD-Violence dataset, due to diminished inter-class separation and increased overlap between normal and abnormal representations. Most notably, eliminating the entailment loss causes a clear degradation to 88.91% AUC on the UCF-Crime dataset and 84.72% AP on the XD-Violence dataset, highlighting its critical role in modeling asymmetric relational constraints within the embedding space, without relying on explicitly defined category-level hierarchies, thereby improving discriminability between semantically similar anomaly patterns. Unlike similarity-based objectives, the entailment formulation imposes a directional constraint in the hyperbolic embedding space that enforces asymmetric relationships between instances, which improves the ranking of ambiguous segments under weak supervision. In its absence, the embedding space degenerates into a purely similarity-driven structure, thereby reducing its ability to separate visually similar normal and anomalous segments. The complete formulation achieves the best performance of 90.62% AUC on the UCF-Crime dataset and 86.93% AP on the XD-Violence dataset, demonstrating that the proposed losses contribute in a complementary manner to jointly enable a discriminative and geometry-aware representation that improves separation and ranking of anomaly patterns for robust anomaly detection. Table 14Ablation study on loss components for UCF-Crime and XD-Violence datasets.ModelUCF-CrimeXD-ViolenceAUC (%)AP (%) only✓×××84.3779.92w/o ✓×✓✓86.1281.45w/o ✓✓✓×89.0585.21w/o ✓✓×✓88.9184.72Complete Model✓✓✓✓90.6286.93 Ablation study on loss components for UCF-Crime and XD-Violence datasets. Limitations and future work Although the proposed framework achieves strong performance across both the UCF-Crime and XD-Violence datasets, certain challenging scenarios still lead to inaccurate anomaly localization and unstable anomaly confidence estimation. To further analyze the behavior of the proposed framework under difficult surveillance conditions, we conduct an error taxonomy analysis on incorrectly localized or misclassified samples from the evaluation set. The failure samples are manually inspected and categorized according to their dominant observable prediction characteristics, and the frequency of each category is estimated from the proportion of samples exhibiting the corresponding failure pattern. The dominant error categories observed in the proposed framework are summarized in Table 15. Table 15Error taxonomy analysis of the proposed PoinCLIP-VAD framework under challenging surveillance scenarios.Error CategoryFrequency (%)Observed Failure BehaviorContext Confusion31.4Normal activities receiving high anomaly confidenceSubtle-Motion Misses26.1Weak anomalies producing low anomaly activationBoundary Drift22.3Imprecise localization near temporal transitionsCrowded-Scene Ambiguity20.2Unstable ranking under dense interactions and occlusion Error taxonomy analysis of the proposed PoinCLIP-VAD framework under challenging surveillance scenarios. These analyses show that context confusion remains the most dominant failure category, particularly in scenarios where normal activities exhibit motion dynamics visually similar to anomalous events, such as rapid crowd movement during sports activities or public gatherings that resemble panic or violent behavior. Another common challenge is subtle-motion misses, which generally occur when anomalous activities occupy only a brief temporal duration or exhibit weak motion intensity, as observed in distant physical altercations or slow object theft, where the anomaly signal remains insufficiently prominent during temporal aggregation. Boundary drift mainly arises near gradual transitions between normal and abnormal activities, especially when seemingly normal interactions progressively escalate into violent behavior, making precise anomaly boundary localization difficult under weak supervision. Similarly, crowded scenes containing overlapping interactions, complex motion trajectories, and severe occlusion, particularly in densely populated street environments or protest scenarios involving multiple simultaneous activities, can further reduce the stability of anomaly ranking by introducing noisy and highly entangled temporal representations. Future research may further investigate uncertainty-aware hyperbolic representation learning for improving anomaly discrimination in semantically ambiguous surveillance scenarios, where uncertainty estimation can help reduce unstable anomaly confidence prediction and improve the reliability of cross-modal alignment. Another promising direction is the integration of audio-visual hyperbolic alignment, which may enable the framework to jointly model complementary acoustic and visual anomaly cues for more robust anomaly understanding in complex real-world environments. In addition, extending the proposed framework toward hierarchical hyperbolic semantic reasoning could enable more effective modeling of multi-scale anomaly relationships and complex contextual dependencies across long-duration surveillance videos. Conclusion This paper introduces PoinCLIP-VAD, an advanced framework for weakly supervised video anomaly detection that integrates the representational power of vision-language models with the geometric expressiveness of hyperbolic space. By embedding both visual and textual features within the Poincaré ball manifold, the model captures latent semantic organization and asymmetric relationships in the embedding space, which traditional Euclidean approaches struggle to represent. The proposed dual-branch architecture, comprising a classification block and a visual-text alignment block, enables effective anomaly detection at both coarse and fine-grained levels. The inclusion of learnable multimodal prompts and cross-space interaction mechanisms ensures curvature-consistent alignment between video and text embeddings under weak supervision, leading to more stable MIL-based ranking and improved discrimination between normal and anomalous events. Extensive evaluations on benchmark datasets demonstrate that PoinCLIP-VAD outperforms existing methods, achieving state-of-the-art performance. Acknowledgements The authors express their sincere gratitude towards their parental organization for providing a research environment. Author contributions D.P.S: Conceptualization, Implementation of methodology, Testing, and validation, Writing - original draft. S.K.P: Conceptualization, Investigation, Supervision, Writing - review & editing. S.K.B: Conceptualization, Investigation, Supervision, Writing - review & editing. P.P.D: Conceptualization, Formal analysis, Investigation, Supervision, Writing - review & editing. R.S.S: Supervision, Writing - review & editing. Funding Open access funding provided by Symbiosis International (Deemed University). Not Applicable Data availability All the Experiments are performed on publicly available datasets. Declarations Competing interest The authors declare no competing interest. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Santos Kumar Baliarsingh, Email: santos.baliarsinghfcs@kiit.ac.in. Ramasubbareddy Somula, Email: ramasubbareddy@sithyd.siu.edu.in. References Associated Data Data Availability Statement All the Experiments are performed on publicly available datasets.