🤖 AI Summary
This work addresses the limitation of existing video anomaly detection methods, which predominantly rely on reactive multiple instance learning and struggle to proactively anticipate anomalies. The authors propose PULS, a continuous semantic world model that maps physical motion representations to a semantic embedding space via a KSD bridge and incorporates an ASP module to enhance the semantic separability of future states—enabling anomaly prediction without multiple instance learning. The study validates the “latent clarity” hypothesis, demonstrating that anticipated future representations exhibit greater semantic separability than current observations and reside on distinct submanifolds. Built upon V-JEPA 2, Qwen3-VL-Embedding-2B, and an L1-surprise gating mechanism, PULS achieves 0.8994 AUROC on UCF-Crime, 0.8162 zero-shot transfer performance on XD-Violence, an 8.9% early-prediction advantage at T−0.5s, and a zero-shot VQA accuracy of 44.5%, substantially outperforming observation-based baselines.
📝 Abstract
Continuous video anomaly detection is dominated by reactive Multiple Instance Learning (MIL) that collapses spatiotemporal features into scalar scores. We introduce PULS (Predictive Unified Latent Space), a continuous semantic world-model pipeline comprising two modules: a 490M-parameter KSD Bridge (Kinematic-to-Semantic Distillation) and a 16.8M-parameter Anticipatory State Predictor (ASP). The KSD Bridge maps V-JEPA 2 physical tensors into the 2048-d Qwen3-VL-Embedding-2B text-aligned hypersphere, trained on a subset of UCF-Crime. This translation alone yields a chunk-level AUROC of 0.8994 for UCF-Crime and 0.8162 for out-of-distribution XD-Violence without MIL or hierarchical fusion.
We introduce and validate the Latent Clarity Hypothesis: because JEPA's temporal predictor discards aleatoric pixel noise while preserving kinematics, anticipated future representations are more semantically separable than observed presents. The ASP sharpens these anticipated future latents, achieving 44.5% mean 14-way zero-shot VQA accuracy (exceeding observation baseline by +9.6 pp). Applying the ASP to Observation Tensors collapses accuracy to 7.3% (random chance), proving Anticipation and Observation occupy distinct sub-manifolds. A Triple-Track Lead-Time protocol with an L1-surprise gate yields a peak +8.9 pp anticipatory advantage at T-0.5s (p < 0.001, N = 1,000 permutation), separating physical anticipation from static scene priors. Zero-shot transfer to XD-Violence confirms that Newtonian-invariant kinematic representations generalize out-of-distribution.