Large Language Models Generate Better Planning Heuristics Without Domain Expertise

Researchers have introduced a planning method that leverages large language models to generate heuristics—evaluation functions that guide search algorithms—directly from problem specifications, eliminating the need for manually engineered domain knowledge.

The Core Innovation#

The technique works by feeding problem definitions to LLMs in the form of successor generators (functions that produce next states), goal tests, and initial conditions written in standard programming code. The models then synthesize custom heuristic functions tailored to each planning task. These heuristics are compiled and fed into classical search algorithms like greedy best-first search, which use them to explore the solution space more efficiently.

Performance and Scope#

Testing across established planning benchmarks revealed that the approach matches or exceeds the performance of existing methods. A key advantage is its flexibility: the method can tackle problems that are cumbersome or impossible to express in traditional planning formalisms, such as those involving intricate numeric constraints or domain-specific state transition mechanics.

Why It Matters#

By automating heuristic design, this work sidesteps the traditional bottleneck of requiring expert knowledge for each new domain. The extensive empirical evaluation characterizes where the approach excels and where it faces limitations, providing practical guidance for practitioners seeking to apply LLM-driven planning to real-world scenarios.