LLMs and Search in Feedback Space for Planning Domain Generation

A new computer science paper investigates using large language models (LLMs) and reasoning models to generate planning domains from natural language descriptions. The research proposes an agentic language model feedback framework that augments descriptions with minimal symbolic information. It evaluates domain quality under symbolic feedback mechanisms including landmarks and output from the VAL plan validator. The study experiments with heuristic search over model space to optimize domain quality. The work addresses the ongoing challenge of producing high-quality planning domains deployable in practice.