arXiv:2605.11723v2 Announce Type: replace 
Abstract: In this paper, we propose Concentrate and Concentrate (CaC), a coarse-to-fine anomaly reward model based on Vision-Language Models. During inference, it first conducts a global temporal scan to anchor anomalous time windows, then performs fine-grained spatial grounding within the localized interval, and finally derives robust judgments via structured spatiotemporal Chain-of-Thought reasoning. To equip the model with these capabilities, we construct the first large-scale generated video anomaly dataset with per-frame bounding-box annotations, temporal anomaly windows, and fine-grained attribution labels. Building on this dataset, we design a three-stage progressive training paradigm. The model initially learns spatial and temporal anchoring through single- and multi-frame supervised fine-tuning, and then is optimized by a reinforcement learning strategy based on two-turn Group Relative Policy Optimization (GRPO). Beyond conventional accuracy rewards, we introduce Temporal and Spatial IoU rewards to supervise the intermediate localization process, effectively guiding the model toward more grounded and interpretable spatiotemporal reasoning. Extensive experiments demonstrate that CaC can stably concentrate on subtle anomalies, achieving a 25.7% accuracy improvement on fine-grained anomaly benchmarks and, when used as a reward signal, CaC reduces generated-video anomalies by 11.7% while improving overall video quality.

يقدم مقال جديد نموذج Concentrate and Concentrate (CaC) الذي يعد نموذج مكافأة لتمييز الشذوذ في الفيديو، والذي يستخدم نماذج الرؤية واللغة لتحسين الكشف عن الشذوذ من خلال نهج هرمي زمني مكاني. يقوم النموذج بإجراء مسح زمني عالمي لتحديد النوافذ الزمنية الشاذة، يليه تثبيت مكاني دقيق واستدلال منظم. تستند هذه الطريقة المبتكرة إلى مجموعة بيانات كبيرة للشذوذ في الفيديو مع تعليقات تفصيلية.

Un nuevo artículo presenta Concentrate and Concentrate (CaC), un modelo de recompensa para la anomalía de video que utiliza Modelos de Visión-Lenguaje para mejorar la detección de anomalías a través de un enfoque jerárquico espaciotemporal. El modelo realiza un escaneo temporal global para identificar ventanas de tiempo anómalas, seguido de un anclaje espacial detallado y razonamiento estructurado. Este método innovador se apoya en un conjunto de datos de anomalías de video a gran escala con anotaciones detalladas.

Un nouvel article présente Concentrate and Concentrate (CaC), un modèle de récompense pour l'anomalie vidéo qui utilise des modèles de vision-langage pour améliorer la détection des anomalies grâce à une approche spatiotemporelle hiérarchique. Le modèle effectue un scan temporel global pour identifier les fenêtres temporelles anormales, suivi d'un ancrage spatial détaillé et d'un raisonnement structuré. Cette méthode innovante est soutenue par un ensemble de données d'anomalies vidéo à grande échelle avec des annotations détaillées.

A new paper introduces Concentrate and Concentrate (CaC), a video anomaly reward model that leverages Vision-Language Models to enhance anomaly detection through a hierarchical spatiotemporal approach. The model conducts a global temporal scan to identify anomalous time windows, followed by fine-grained spatial grounding and structured reasoning. This innovative method is supported by a large-scale video anomaly dataset with detailed annotations.