LLM-Evolved Pattern Generators for Optimal Classical Planning

2026-06-01 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

A new method, "LLM-Evolved Pattern Generators," introduces the first approach for learning domain-dependent heuristics that are admissible by design for optimal classical planning. Unlike existing learned heuristics focused on satisficing planning, this technique preserves the optimality guarantees of A* search. It employs an LLM-driven evolutionary program-synthesis framework to generate, for each planning domain, a program that constructs a pattern collection for any task. These patterns are then combined admissibly using saturated cost partitioning. Empirically, the learned programs provide interpretable domain-specific insights, operate with negligible overhead at test time, and achieve coverage matching state-of-the-art domain-independent baselines across several domains, while evaluating each state substantially faster.

Key takeaway

For research scientists developing optimal classical planning systems, this method offers a new path to admissible heuristics. You should consider integrating LLM-driven program synthesis to generate domain-dependent pattern collections, ensuring A* search optimality while significantly accelerating state evaluation. This approach provides interpretable insights and matches state-of-the-art coverage, potentially streamlining your heuristic development workflow.

Key insights

LLM-driven program synthesis generates admissible, domain-dependent heuristics for optimal classical planning, preserving A* search optimality.

Principles

• Admissibility is crucial for optimal planning heuristics.
• LLMs can synthesize programs for heuristic generation.
• Saturated cost partitioning ensures pattern admissibility.

Method

An LLM-driven evolutionary program-synthesis framework generates domain-specific programs. These programs produce pattern collections for planning tasks, which are then combined admissibly via saturated cost partitioning to form heuristics.

In practice

• Apply LLM-synthesized heuristics for optimal A* planning.
• Use generated patterns for faster state evaluation.
• Gain interpretable domain-specific insights.

Topics

• Optimal Classical Planning
• Admissible Heuristics
• LLM-driven Program Synthesis
• Pattern Generators
• A* Search
• Saturated Cost Partitioning

Best for: AI Scientist, Research Scientist

Related on AIssential

• From Monolithic Models to Decision Ecosystems: Orchestrating Optimization with AI Precision Agents — Operational AI's future lies in governed ecosystems of specialized agents, not monolithic models.
• Two-Fidelity Best-Action Identification for Stochastic Minimax Tree — 2FFS is a two-fidelity tree-search algorithm for stochastic minimax trees, balancing cheap, biased heuristics with expensive, accurate rollouts.
• Harness-1: Reinforcement Learning for Search Agents with State-Externalizing Harnesses — Harness-1 improves RL search agents by offloading routine state management to an external harness, boosting recall and generalization.
• COMAP: Co-Evolving World Models and Agent Policies for LLM Agents — COMAP co-evolves world models and agent policies through closed-loop interaction for adaptive LLM agents.
• SIRI: Self-Internalizing Reinforcement Learning with Intrinsic Skills for LLM Agent Training — SIRI enables LLM agents to autonomously discover and internalize skills, reducing external dependencies and improving long-horizon task performance.
• Coordination Graphs for Constrained Multi-Agent Reinforcement Learning — The CG-CMARL framework efficiently scales multi-agent reinforcement learning by combining coordination graphs and Lagrangian duality for constraint handling.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.