LLM-Guided Evolutionary Search Reformulates Constraint Models for Faster Solving

arXiv·low signalXBlueskyLinkedInCopy link

Recent work established that LLMs can produce correct declarative constraint models from natural language, but correctness is not the whole problem — solver performance remains highly sensitive to modelling choices, so a correct model can still be prohibitively expensive to solve. This work investigates whether LLMs can automate the reformulation step, using evolutionary search guided by the model to find equivalent formulations that compile to faster solver runs. It targets a real gap in the natural-language-to-solver pipeline that shows up across scheduling, planning, and configuration workloads in industry.