A Physics-Guided and Self-Adaptive Multi-Agent Framework for Jet Anomaly Detection

Jet anomaly detection in a high-energy physics is a non-stationary task that is fuelled by shifts in the domain due to pile-up and predominantly by background noise, and dynamically changing relationships between jet constituents in such a scenario, where a conventional graph neural network architecture is frequently inadequate in terms of robustness and interpretability. Physics-Self-Adaptive Multi-Agent System (PhySA-MAS) is a physics-directed, self-adaptive multi-agent architecture that proposes jet analysis as a decentralized and dynamically reconfigurable reasoning scheme. It does not use one monolithic model but instead integrates specialist agents dealing with meta-learning, relational reasoning, communication and topology control which can vary their interactions depending on event-level physics. The energy conservation constraints are incorporated in graph message passing to provide physical consistency, and a reinforcement driven topology controller rewires inter-agent communication dynamically to forces according to anomalous patterns. An additional communication strategy, anchor-peer communication, ensures the further stabilization of learning through the reduction of gradient conflict and the amplification of the signals related to anomalies, which, in combination, offers a powerful and structurally understandable alternative to fixed deep learning models.