Project Description

Recent works like xGrammar have introduced constrained decoding which guarantees that LLM decode outputs that conform to user-defined constraints (e.g., extract date of birth from ID card in a DDMMYYYY format). Constrained decoding yields meaningful accuracy improvements over unconstrained decoding: rather than hoping that the model gets it right, it makes incorrect outputs structurally impossible. This project aims to extend that foundation in two significant directions. First, we move beyond syntactic constraints to semantic constraints. For example, in addition to ensuring that dates are well-formed, enforce relationships like the date of birth precedes the date of issue by at least 18 years. This opens up a richer class of correctness guarantees. Second, we explore constrained decoding in a broader set of applications, such as in document processing pipelines, where structured extraction from noisy, varied inputs stands to benefit greatly from hard output guarantees, and in SMT solving, where the correctness of generated expressions is non-negotiable.