LLM-Evolved Domain-Independent Heuristics for Symbolic AI Planning

Researchers used large language models and evolutionary search to create the first domain-independent heuristics for symbolic AI planning that surpass hand-engineered baselines. These evolved heuristics, written in C++, solve more planning tasks than existing state-of-the-art approaches and maintain the soundness guarantees of traditional planners.

This research represents a significant milestone in automated algorithm design, demonstrating that LLMs coupled with evolutionary techniques can discover generalizable solutions that exceed decades of human expert work. The breakthrough lies not in domain-specific optimization but in producing heuristics that function across arbitrary planning problems—a substantially harder challenge that requires discovering fundamental principles rather than task-specific tricks.

The methodology is noteworthy for its pragmatic approach to LLM limitations. Rather than expecting language models to generate perfect heuristics directly, the researchers treat LLM-generated code as mutation candidates within an evolutionary framework, using a MAP-Elites archive to balance the speed-informedness tradeoff. This hybrid approach sidesteps the tendency of LLMs to hallucinate or produce suboptimal solutions by allowing natural selection to filter candidates.

The findings challenge conventional wisdom in multiple ways. Counterintuitively, seeding evolution from trivial blind heuristics outperformed starting from sophisticated hand-engineered baselines like FF, suggesting that evolutionary processes benefit from simpler starting points. Additionally, the analysis reveals that LLM reasoning effort primarily affects code compilation success rather than solution quality—a distinction that could redirect how researchers structure such experiments.

For the broader AI community, this work validates that learned algorithms can penetrate traditionally human-expert-dominated domains. The immediate impact extends to planning systems used in robotics, scheduling, and automated reasoning. However, the results remain confined to symbolic planning; broader implications for other algorithmic domains remain unexplored. The integration of these heuristics as drop-in replacements preserves existing system guarantees, reducing adoption friction for practitioners.

• →LLM-evolved heuristics now exceed hand-engineered symbolic AI planning algorithms on unseen domains
• →Evolutionary search combined with LLM mutation produces generalizable solutions across arbitrary planning tasks
• →Seeding evolution from simple baselines unexpectedly outperforms initialization from sophisticated expert-designed heuristics
• →Evolved heuristics maintain soundness and completeness guarantees as C++ drop-in replacements for existing planners
• →LLM reasoning effort predominantly affects code compilation rather than solution quality in evolutionary contexts