MIT Proposes PDDL-INSTRUCT, Boosting Llama-3-8B Planning Accuracy to 94%

MIT researchers recently introduced a new method called PDDL-INSTRUCT, designed to enable large language models (LLMs) to perform genuine step-by-step logical reasoning in planning problems, rather than relying on superficial pattern matching. This research has drawn significant attention because it demonstrates a profound leap in model planning capabilities, rather than just an incremental improvement.

Key Details and Verification Mechanism

According to a thread by @mdancho84, the core of this method lies in a two-step process of training and inference. First, the LLM is trained using data containing both correct and incorrect plans alongside their explanations. Second, during actual model inference, an external verification mechanism is introduced. This external process checks each generated reasoning step to ensure logical reliability. The researchers emphasize that this approach drives the model to execute stronger logical reasoning checks.

Performance

In terms of concrete results, the method delivered a dramatic performance boost. On a specific planning benchmark, the Llama-3-8B model's accuracy surged from an initial 28% directly to 94%. The author notes that a leap of this magnitude represents a paradigm shift in model capability, proving the immense potential of this approach in enhancing LLM planning and reasoning skills.

2026-07-09 ~ 2026-07-09 · 8 related posts