MIT Fixes AI Planning: 28% to 94%
MIT just dropped the paper that fixes AI planning. 28% → 94% accuracy by teaching LLMs to verify each step instead of bullshitting. They call it PDDL-INSTRUCT. The secret sauce? Making models check preconditions/effects like autistic savants instead of creative writers. Every "AI planning" demo you've seen was fake until now.
Published September 24, 2025