Distill-Belief: Closed-Loop Inverse Source Localization and Characterization in Physical Fields

Distill-Belief (arXiv:2604.26095) introduces a teacher-student framework for closed-loop inverse source localization, decoupling Bayesian correctness from inference speed. A particle-filter teacher provides dense information-gain supervision while a compact student distills posterior statistics for real-time control, achieving constant per-step cost at deployment. Evaluated across seven field modalities, this directly addresses reward hacking in belief-space RL where fast surrogate models exploit approximation errors. Roboticists and sensor-fusion engineers tackling radiation, chemical, or acoustic source localization under time constraints should prioritize this over standard POMDP or GP-based active sensing baselines.