Real-world decision problems often combine two kinds of constraints: hard constraints that must be satisfied, and soft constraints that represent preferences and can be violated. Translating such problems into formal optimization models is a challenging task, especially when the input is described in natural language.
The paper introduces ASTOR, a benchmark and method for autoformulation of soft-constraint optimization. The work leverages a formal grammar to structure the translation process, aiming to improve how systems convert problem descriptions into solvable models.
By providing a standardized benchmark, ASTOR offers a way to evaluate and compare different autoformulation approaches. This is a step toward more reliable automated decision-support tools that can handle the nuance of soft constraints in practical settings.