CP-SynC: Multi-Agent Zero-Shot Constraint Modeling in MiniZinc with Synthesized Checkers
CP-SynC (Constraint Programming modeling with Synthesized Checkers) represents a collaborative framework for zero-shot constraint modeling within MiniZinc. This system orchestrates modeling agents tasked with creating and refining potential models, alongside validation agents that develop semantic checkers to assess correctness. To reduce the noise typically found in individual LLM outputs, CP-SynC investigates various modeling paths simultaneously and utilizes selection agents for final model determination through multi-agent evidence aggregation. This method tackles the difficulty of converting natural language problem statements into functional constraint programming models, an area where LLMs frequently encounter subtle semantic inaccuracies without oracle validation during testing.
Key facts
- CP-SynC is a multi-agent workflow for zero-shot constraint modeling in MiniZinc.
- It coordinates modeling agents and validation agents.
- Validation agents synthesize semantic checkers to provide feedback.
- Multiple modeling trajectories are explored in parallel.
- Selection agents choose the final model via multi-agent evidence aggregation.
- The approach addresses LLM struggles with semantic errors in constraint modeling.
- The paper is available on arXiv with ID 2605.01675.
- The announcement type is new.
Entities
Institutions
- arXiv