formal-abstractions

The paper introduces SDDL, a neuro-symbolic framework that improves combinatorial optimization accuracy in resource-constrained language models by translating natural-language problems into formal representations, resulting in higher feasibility rates compared to direct-generation and solver-code baselines.