Minimal-Core-Guided Repair for Neuro-Symbolic Constraint Solving

A recent preprint on arXiv (2608.14771) introduces a technique aimed at enhancing the dependability of language models in addressing constraint issues by correcting inaccurate translations. This method substitutes error notifications with proofs; if a produced program is unsatisfiable, it identifies a minimal unsatisfiable core based on the model's constraints, offering a clear signal that pinpoints the error. In testing across 77 problems with a precise oracle, the translation to Answer Set Programming proved faithful in six out of seven domains, with the sole exception being aggregate coverage scheduling.