Auxiliary control variables such as gating, routing, and reliance coefficients are widely used to arbitrate between competing information sources in neural systems, yet they are typically optimized only indirectly through downstream task losses. As a result, these variables can collapse into low-sensitivity, low-variance, or behaviorally inconsistent regimes while nominal predictive performance remains high. We introduce function-space behavioral priors that preserve the functional role of a reliance variable α by constraining the local sensitivity of the decision function with respect to α. An effect prior prevents functional suppression through a reliability-weighted sensitivity floor, while a monotonicity prior promotes directional consistency with internal arbitration signals. We formulate training as a hybrid MAP objective combining task likelihood, Bayesian regularization, and derivative-based behavioral constraints, and provide local guarantees showing that the priors penalize collapse-prone regimes. Across two domains, perturbation settings, baseline comparisons, and ablations, the priors improve alignment between α and retrieval reliability, stabilize control dynamics, and improve robustness and predictive performance under higher distributional complexity. These results suggest that preserving behavioral coherence in internal control variables can serve as a principled inductive bias for reliability-aware neural decision systems.

Menu

Ms. Mayra Bornacelly