LatticeBridge: New inference method boosts structured sequence faithfulness

Structured sequence generation—think data-to-text or constrained summarization—often requires a model to satisfy multiple input-derived constraints in a single output. Standard decoding methods (greedy, beam search) assign high probability to fluent continuations but place low mass on continuations that realize all required anchors jointly. LatticeBridge reframes this as a rare-event sequential inference problem.

The framework combines a compact prefix language model with instance-compiled surface automata (constraint representations built per input, not from curated classes) and a twisted sequential Monte Carlo (SMC) decoder that uses resampling, multilevel splitting, and a source-support proposal term derived from input phrases. On 2,610 validation tasks spanning CommonGen (generating sentences from concepts), E2E NLG (restaurant descriptions), and WikiBio (biography generation), the particle decoder achieves significantly higher exact anchor satisfaction and mean anchor coverage than greedy, beam-filtered, and best-of-k ancestral baselines—all under a shared proposal model.

The evaluation goes beyond simple accuracy: it reports required-anchor coverage, source coverage, source-intrusion diagnostics, overlap, runtime, and particle statistics jointly. This comprehensive benchmark characterizes the faithfulness-overlap-latency frontier, showing that LatticeBridge delivers more reliable constraint enforcement without sacrificing fluency. The code and benchmark files are publicly available.