NBSR: Neural Bayesian Sequential Routing for uncertainty-aware AI inference

A new research paper from Yongchao Huang presents Neural Bayesian Sequential Routing (NBSR), a mathematically grounded framework that replaces static neural network forward passes with active, uncertainty-aware evidence accumulation. NBSR operates over a hierarchical Directed Acyclic Graph (DAG) of neural experts, each querying a persistent global knowledge oracle. The oracle returns positive evidence vectors that act as pseudo-counts within a Dirichlet–Categorical conjugate model. Beliefs are updated via exact conjugate addition, and a Gumbel-Softmax Straight-Through estimator enables discrete, path-dependent routing decisions while preserving differentiable gradients for end-to-end training.

Key theoretical results are proven: under strictly positive evidence extraction, total Dirichlet precision increases monotonically along any valid trajectory, and marginal predictive variance is bounded – formalizing a process the author calls “hypothesis sharpening.” Under idealized capacity and optimization assumptions, the terminal Dirichlet expectation recovers the Bayes-optimal conditional distribution. These properties give NBSR natural mechanisms for uncertainty quantification, entropy-based early exiting, out-of-distribution (OOD) abstention, and cost-aware evidence acquisition.

Empirical evaluations span visual categorization, structured medical diagnosis, language modeling, partially observable control, and Bayesian experimental design. NBSR achieves competitive predictive performance while offering transparent routing traces, path-dependent evidence attribution, and uncertainty-aware decision control. The framework directly addresses growing demands for interpretable, modular, and resource-rational AI agents – a step toward neural systems that know when to think harder, when to stop, and when to abstain.