LLM-generated reasoning graphs, referred to as mission-specific graphs (MSGs), are increasingly used for video anomaly detection (VAD) and recognition (VAR). However, they are typically treated as fixed despite being generic and distribution-deficient. Conventional graph structure refinement (GSR) methods are ill-suited to this setting, as they rely on learning structural distributions that are absent in LLM-generated graphs. We propose HDC-constrained Graph Structure Refinement (HDC-GSR), a new paradigm that directly optimizes a decodable, task-aligned graph representation in a single hyperdimensional space without distribution modeling. Leveraging Hyperdimensional Computing (HDC), our framework encodes graphs via binding and bundling operations, aligns the resulting graph code with downstream loss, and decodes edge contributions to refine the structure. We instantiate this approach as MissionHD for weakly supervised VAD/VAR and demonstrate consistent performance gains on benchmark datasets.

翻译：LLM生成的推理图（称为任务特定图，MSG）越来越多地用于视频异常检测（VAD）与识别（VAR）。然而，这些图尽管具有通用性，却普遍存在分布缺失问题，且通常被当作固定结构处理。传统的图结构精化（GSR）方法依赖学习结构分布，而LLM生成的图中缺乏此类分布，因此不适用于此场景。我们提出HDC约束的图结构精化（HDC-GSR），这是一种无需分布建模、直接在单一超维度空间中优化可解码且任务对齐的图表示的新范式。该框架利用超维度计算（HDC），通过绑定与聚合操作对图进行编码，将生成的图代码与下游损失对齐，并通过解码边贡献度来精化结构。我们将该方法实例化为MissionHD，用于弱监督的VAD/VAR任务，并在基准数据集上验证了其持续的性能提升。