Entropy-Induced Collapse in Learned Decision Policies under Partial Observability and Architectural Mitigation

ACM/VERSION, This paper studies a previously undercharacterized failure mode of learned decision policies operating under partial observability. While data-driven policies often demonstrate superior efficiency in low-noise environments, we show that their operational stability degrades non-linearly as observation uncertainty increases. Rather than exhibiting graceful performance decay, learned policies undergo a sharp collapse characterized by high-variance behavior, excessive action switching, and elevated risk exposure. We formalize this phenomenon as entropy-induced policy collapse and analyze it within a Partially Observable Markov Decision Process (POMDP) framework. Using entropy-based indicators derived from policy outputs, we identify a narrow transition region in which multiple instability signals co-occur, revealing an information-theoretic boundary on reliable policy operation. Through extensive simulation experiments, we demonstrate that this collapse is consistent across multiple model classes and is not resolved by architectural complexity alone. To address this limitation, we propose a Hybrid Meta-Control (HMC) architecture that supervises learned policies with entropy-aware fallback mechanisms. The proposed approach dynamically arbitrates between learned and heuristic controllers based on real-time estimates of uncertainty and instability. Our results show that the HMC architecture reduces collapse frequency by over 90% while preserving most of the efficiency gains of learned policies. These findings suggest that robustness in autonomous decision systems is fundamentally an architectural property rather than a purely statistical one, with implications for the deployment of learned controllers in safety- and reliability-critical domains.