Understanding anomalies in real-world videos requires reasoning about when events occur, where they happen, and what actions unfold. Existing benchmarks, however, largely ignore one or more of these dimensions, making it difficult to evaluate whether multimodal large models (MLLMs) truly understand abnormal events beyond simple detection. To address this gap, we introduce Anom3W, a benchmark designed to evaluate MLLMs across temporal, spatial, and semantic dimensions. Our semi-automatic pipeline consolidates heterogeneous anomaly resources into verified spatio-temporal annotations, enabling rigorous, reasoning-oriented evaluation. By treating spatial boxes and tubes as evidence cues, Anom3W challenges models to reason about entities, locations, and behaviors, rather than merely flag anomalies. Experiments demonstrate that even state-of-the-art MLLMs struggle, highlighting the critical importance of spatial grounding and semantic understanding. Anom3W establishes a new standard for interpretable, context-aware video anomaly reasoning.

Vision-Language-Action (VLA) models have shown promising progress in language-conditioned robotic manipulation. However, existing datasets and benchmarks mainly evaluate task completion under predefined settings, offering limited insight into model reasoning under increasing spatial and procedural complexity. We introduce RoboSPA (Robot Spatial-Procedural Assessment), a large-scale robotic manipulation dataset and benchmark for diagnosing embodied reasoning in VLA models. RoboSPA focuses on two core dimensions, Fine-Grained Spatial Reasoning and Long-Horizon Procedural Planning, covering 10 task categories and 56 base tasks. Each task is instantiated across five difficulty levels, yielding 280 variants with increasing spatial ambiguity and procedural complexity. We collect 527K trajectories across multiple embodiments and diverse scenes. Beyond binary success rate, RoboSPA introduces diagnostic metrics for more detailed evaluation. Experiments on representative VLA models show that current systems still struggle with complex spatial relations, precise low-level execution, and memory-intensive planning. These results establish RoboSPA as a challenging diagnostic benchmark for developing more capable, reliable, and generalizable embodied agents. Our data and code are available at https://github.com/fanzhenxuan/RoboSPA.