CaC：通过分层时空聚焦推进视频奖励模型
CaC: Advancing Video Reward Models via Hierarchical Spatiotemporal Concentrating
论文提出 Concentrate and Concentrate（CaC），一种基于 Vision-Language Models 的由粗到细异常奖励模型，先全局时间扫描锚定异常时间窗，再在区间内做细粒度空间定位，最后通过结构化时空 Chain-of-Thought 推理给出判断。
Authors:Jiyuan Wang, Huan Ouyang, Jiuzhou Lin, Chunyu Lin, Dewen Fan, Boheng Zhang, Haonan Fan, Honglie Wang, Yiyang Fan, Zhenlong Yuan, Yongrui Heng, Guosheng Lin, Fan Yang
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.
| Comments: | Accept to NeurIPS 2026, 27 pages, 10 figures |
| Subjects: | Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2605.11723 [cs.CV] |
| (or arXiv:2605.11723v4 [cs.CV] for this version) | |
| https://doi.org/10.48550/arXiv.2605.11723 arXiv-issued DOI via DataCite |
Submission history
From: JiYuan Wang [view email]
[v1]
Tue, 12 May 2026 08:08:33 UTC (7,115 KB)
[v2]
Thu, 28 May 2026 12:50:43 UTC (7,116 KB)
[v3]
Fri, 25 Sep 2026 05:22:14 UTC (7,116 KB)
[v4]
Thu, 29 Oct 2026 14:19:19 UTC (7,124 KB)
来源：arXiv cs.AI · arxiv.org