LLMs Generate Streamliner Constraints for Answer Set Programming, Achieving 4-5x Speedups
A novel research methodology utilizes Large Language Models (LLMs) to create streamlined constraints for Answer Set Programming (ASP), thereby improving knowledge representation and reasoning capabilities. This technique employs LLMs alongside ASP encoding and minimal training examples to suggest candidate constraints that narrow the search space for combinatorial challenges. Constraints that introduce syntax errors or negatively impact performance are eliminated. The remaining streamliners are assessed against the initial encoding, with a virtual best encoding (VBE) yielding speed enhancements of 4-5 times across three benchmarks from the ASP Competition: Partner Units Problem, Sokoban, and Towers of Hanoi. This study, published on arXiv (identifier 2604.19251v1), showcases notable performance gains via AI-driven constraint refinement.
Key facts
- Streamliner constraints reduce combinatorial problem search space by eliminating solution portions
- LLMs generate candidate constraints for Answer Set Programming given encoding and training instances
- Candidates causing syntax errors or making satisfiable instances unsatisfiable are discarded
- Virtual best encoding selects fastest option between original and streamlined variants per instance
- Achieves 4-5x speedups on Partner Units Problem, Sokoban, and Towers of Hanoi benchmarks
- Adapts StreamLLM approach from Constraint Programming to Answer Set Programming
- Research announced on arXiv with identifier 2604.19251v1
- Multiple LLMs produce varying candidate constraints
Entities
Institutions
- arXiv