A0376
Title: Belief distortion under stress using trust decay and kl-regularized inference
Authors: Gabriel Nixon - New York University (United States) [presenting]
Abstract: When agents operate under stress, they tend to overreact, ignore prior beliefs, or misallocate trust. A general framework is proposed to model this using kl-regularized updates with entropy limits and decaying trust. The idea is to simulate how beliefs shift when information is incomplete, reliability degrades, or environments change. The method introduces a budgeted form of belief distortion where each update trades off staying close to prior assumptions with adapting to new shocks. Trust decay is modeled as a weighted penalty that gradually reduces the influence of old beliefs. Diagnostics are included to track regret, drawdowns, and entropy flow over time. Applications span any sequential decision setup under uncertainty, including finance, policy simulations, or dynamic systems. Several open problems are outlined, such as whether belief curvature can flag regime shifts or how repeated interventions distort entropy. The setup supports causal belief adjustments, mirror-descent-style learning, and stress-aware regularization that holds up even under regime breaks. It is designed to be modular, interpretable, and fast enough to use in real-time. The focus is not on perfect predictions but on understanding how agents should bend their beliefs without breaking them when operating under pressure.
