Research1d ago
Distill-Belief Solves Reward Hacking In Source Localization
The Distill-Belief framework uses a particle-filter teacher to provide dense information-gain signals to a compact student model. This decoupling prevents policies from exploiting approximation errors during Bayesian inference. It enables mobile agents to localize physical sources under strict time constraints. Practitioners gain a method to maintain uncertainty accuracy without sacrificing real-time efficiency.