Large Language Model Meets Constraint Propagation

TL;DR

This paper introduces an improved method for constraint-aware text generation by combining Large Language Models with Constraint Programming and Masked Language Models, resulting in more reliable and feasible outputs for constrained tasks.

Contribution

It presents an enhanced GenCP framework that integrates MLMs for bidirectional constraint propagation, bridging token prediction and structured constraint enforcement.

Findings

Improved constraint satisfaction in text generation tasks.

Enhanced performance on COLLIE benchmarks with domain preview.

Trade-offs include increased MLM calls and backtracking.

Abstract

Large Language Models (LLMs) excel at generating fluent text but struggle to enforce external constraints because they generate tokens sequentially without explicit control mechanisms. GenCP addresses this limitation by combining LLM predictions with Constraint Programming (CP) reasoning, formulating text generation as a Constraint Satisfaction Problem (CSP). In this paper, we improve GenCP by integrating Masked Language Models (MLMs) for domain generation, which allows bidirectional constraint propagation that leverages both past and future tokens. This integration bridges the gap between token-level prediction and structured constraint enforcement, leading to more reliable and constraint-aware text generation. Our evaluation on COLLIE benchmarks demonstrates that incorporating domain preview via MLM calls significantly improves GenCP's performance. Although this approach incurs additional MLM calls and, in some cases, increased backtracking, the overall effect is a more efficient use of LLM inferences and an enhanced ability to generate feasible and meaningful solutions, particularly in tasks with strict content constraints.