MIT Researchers Achieve 94% Accuracy in AI Planning Tasks with New PDDL-INSTRUCT Framework
Summary
MIT researchers develop PDDL-INSTRUCT framework that boosts AI planning accuracy to 94% on complex tasks, delivering up to 66% improvement over existing models by combining logical reasoning with step-by-step plan validation.
Key Points
- MIT CSAIL researchers develop PDDL-INSTRUCT, an instruction-tuning framework that improves AI planning by combining logical chain-of-thought reasoning with external plan validation
- The system achieves 94% accuracy on Blocksworld planning tasks using a tuned Llama-3-8B model, representing up to 66% absolute improvement over baseline models
- PDDL-INSTRUCT trains models to explain plan failures and validate each step through external verification, showing dramatic improvements especially on challenging Mystery Blocksworld tasks