Model Space Reasoning as Search in Feedback Space for Planning Domain Generation

A research team from IBM Research and collaborating institutions has introduced a new framework that tackles the persistent challenge of generating usable planning domains from natural language descriptions. While large language models (LLMs) have shown promise in this area, they often fall short of producing the high-quality, logically consistent domains required for real-world deployment in fields like robotics, logistics, and automated reasoning. The paper, "Model Space Reasoning as Search in Feedback Space for Planning Domain Generation," proposes a solution by creating an agentic LLM framework that is augmented with symbolic feedback.

The core innovation lies in its hybrid architecture. The system doesn't rely solely on the LLM's internal reasoning. Instead, it uses the LLM as an agent that proposes domain models. These proposals are then evaluated by external, symbolic tools like the VAL plan validator—a classic AI planner verifier—and landmark analysis, which identifies critical steps in a plan. This feedback is fed back to the LLM agent, guiding it through a heuristic search across the space of possible domain models. The goal is to iteratively refine the domain until it meets rigorous symbolic criteria for correctness and completeness, bridging the gap between flexible natural language understanding and precise, executable formal logic.