LLMs Gain Robust Planning with Symbolic Self-Refinement Framework

Large language models (LLMs) often struggle with complex, multi-step planning, frequently generating plans that are either unfeasible or incorrect. This new research introduces a framework designed to enhance the reliability and robustness of LLM-based planning. The core idea is to integrate symbolic reasoning and iterative self-correction into the LLM's planning process. The framework works by first translating logical symbols into natural language descriptions, helping the LLM better understand task constraints. A crucial component is a symbolic verifier that identifies errors in the LLM's proposed plans and translates these errors into corrective instructions. Additionally, a plan recognizer is used to assess whether the current plan is progressing towards the desired goal, providing further guidance for refinement. Empirical evaluations show that this approach significantly improves both the feasibility and correctness of LLM-generated plans in long-horizon tasks. This advancement suggests a path towards more trustworthy and dependable AI systems capable of handling intricate decision-making scenarios.