MIT Researchers Enhanced Artificial Intelligence (AI) 64x Better at Planning, Achieving 94% Accuracy

MIT CSAIL researchers have developed PDDL-INSTRUCT, an instruction-tuning framework that significantly enhances the logical reasoning and validity of multi-step plans generated by 8-billion-parameter language models like Llama-3-8B. By integrating explicit state and action semantics with external plan validation (VAL), the approach addresses the common issue of LLMs producing plausible but logically invalid plans, achieving up to 94% validity on the Blocksworld benchmark and up to a 66% improvement over previous methods. This innovative method combines logical chain-of-thought prompting with ground-truth plan validation, where