How Far Are LLMs from Symbolic Planners? An NLP-Based Perspective
The paper investigates whether Large Language Models (LLMs) can reliably act as planners by treating planning as a Natural Language Processing (NLP) task. The authors propose a pipeline that evaluates, recovers, and completes flawed LLM plans using NLP and symbolic planners. Findings show LLMs lack underlying reasoning, with only about 2.65% of generated actions being executable, indicating they are far from the reliability of classical symbolic planners.