LatticeBridge: Rare-Event Sequential Inference for Faithful Structured Sequence Synthesis

LatticeBridge introduces a novel approach to structured sequence generation that addresses the challenge of satisfying multiple input-derived constraints simultaneously. It employs a compact prefix language model, instance-compiled surface automata, and a twisted sequential Monte Carlo (SMC) decoder, which enhances exact anchor satisfaction and mean anchor coverage across 2,610 validation tasks from datasets like CommonGen and E2E NLG. This methodology is significant for practitioners as it improves the reliability of model outputs in scenarios requiring adherence to complex constraints, offering a more robust framework for faithful sequence synthesis.