Research21h ago
Distill-Belief Solves Reward Hacking In Source Localization
A new teacher-student framework called Distill-Belief decouples Bayesian correctness from efficiency to locate physical sources. It uses a particle-filter teacher to provide dense information-gain signals, preventing the student model from exploiting approximation errors. This approach stops reward hacking during uncertainty estimation. Practitioners can now deploy faster belief models without sacrificing localization accuracy.