LLM-Evolved Pattern Generators for Optimal Classical Planning

Researchers have developed a novel method using large language models and evolutionary algorithms to automatically generate admissible heuristics for optimal classical planning problems. Unlike existing learned heuristics that improve search speed but cannot guarantee optimal solutions, this approach preserves A* optimality guarantees while matching or exceeding the performance of traditional domain-independent methods.

This research addresses a fundamental limitation in AI planning by bridging the gap between learned heuristics and optimality guarantees. Traditional domain-independent heuristics like LM-Cut provide optimality assurances but lack domain-specific optimization, while recent machine learning approaches sacrifice admissibility for improved search guidance. The proposed LLM-evolved pattern generator framework solves this by having an evolutionary algorithm discover domain-specific abstraction patterns that inherently maintain admissibility constraints, then combines them through saturated cost partitioning.

The method represents a convergence of symbolic AI planning and modern machine learning techniques. By using program synthesis guided by large language models, the system discovers interpretable domain insights rather than opaque neural network representations. This transparency is valuable for understanding why the heuristics perform well and ensures they remain theoretically sound.

The practical implications are significant for planning-dependent applications in robotics, logistics, and automated reasoning. The approach achieves competitive coverage with state-of-the-art baselines while evaluating states substantially faster, suggesting improved real-world performance without sacrificing optimality guarantees. The negligible runtime overhead at test time makes deployment feasible for time-sensitive applications.

Future developments may involve scaling this method to more complex domains, investigating whether learned patterns transfer across related problem classes, and integrating these heuristics into modern planning systems used in industry. The success of this hybrid symbolic-neural approach could inspire similar methods for other combinatorial optimization problems.

• →First method to learn admissible heuristics for optimal planning, preserving A* optimality guarantees while improving performance
• →Uses LLM-driven evolutionary programming to discover domain-specific pattern abstractions rather than direct state-to-value mappings
• →Generated heuristics run with negligible overhead and evaluate states faster than traditional domain-independent baselines
• →Learned programs encode interpretable, human-understandable domain insights instead of black-box neural representations
• →Approach combines multiple patterns through saturated cost partitioning to maintain admissibility constraints