Reasoning is formalized via optimal transport to bound transformers' OOD generalization by architectural Lipschitz continuity and approximation limits, proving depth is needed for backtracking and shift-invariant attention reduces risk.
A neuro-symbolic framework pairs neural graph proposals with symbolic SMT solvers for hard-constraint satisfaction, achieving over 95% in-distribution and 64, 86% zero-shot rule compliance on the MolSAT benchmark.