Neuro-Symbolic Framework Boosts Small LLM Scheduling Accuracy

Combinatorial scheduling problems, which involve finding feasible solutions within vast search spaces while adhering to complex constraints, pose a significant challenge for language models. This difficulty is particularly acute for smaller, resource-constrained models that often struggle to maintain feasibility when generating schedules directly from natural language descriptions. To overcome these limitations, researchers introduced SDDL (Solver-Delegated Declarative Language), a neuro-symbolic framework. SDDL works by translating natural-language scheduling problems into concise, solver-aligned representations of tasks, resources, constraints, and objectives. It then delegates the low-level modeling and search processes to a deterministic compiler and an external solver. Evaluated on a dataset of 300 scheduling problems, SDDL dramatically improved the independently verified feasibility for every resource-constrained model tested. The strongest SDDL configurations achieved feasibility rates of 55.3% and 28.3%, a significant leap from direct-generation baselines of 23.7% and 1.3%, and even outperforming solver-code baselines. This approach allows smaller models to rival the performance of substantially larger frontier models by focusing on problem structure rather than direct solution generation.