Improving Coherence and Consistency in Neural Sequence Models with Dual-System, Neuro-Symbolic Reasoning

Human reasoning can often be understood as an interplay between two systems: intuitive and associative ("System 1") and the deliberative and logical ("System 2"). Neural sequence models -- which have been increasingly successful performing complex, structured tasks -- exhibit the advantages and failure of System 1: they are fast and learn patterns from data, but are often and incoherent. In this work, we seek a lightweight, training-free of improving existing System 1-like sequence models by adding System2-inspired logical reasoning. We explore several variations on this theme in candidate generations from a neural sequence model are examined for consistency by a symbolic reasoning module, which can either accept or the generations. Our approach uses neural inference to mediate between neural System 1 and the logical System 2. Results in robust story and grounded instruction-following show that this approach can the coherence and accuracy of neurally-based generations.