A new procedure-planning method learns to ignore actions that don't make sense for the task at hand.
Researchers propose CEFITO, a method for procedure planning in instructional videos: the task of mapping a sequence of actions from a starting state to a goal state. Existing approaches predict that sequence with feed-forward networks or diffusion-based inference, treating every possible action as equally plausible. CEFITO instead learns an action-conditioned representation space and reformulates planning as a constrained optimization problem, explicitly excluding irrelevant actions during inference rather than after the fact. The paper reports state-of-the-art accuracy on two established procedure-planning benchmarks.
That's a meaningful shift for a subfield where models often propose technically fluent but logically nonsensical steps, like suggesting you fry an egg before cracking it. Constraining the action space up front, instead of hoping a network learns real-world logic implicitly, is a more direct fix than throwing more data or bigger diffusion models at the problem.
It's worth noting the gains come from the authors' own benchmark runs on a preprint that hasn't been peer reviewed, and procedure planning on curated instructional videos is a much cleaner test than the open-ended, ambiguous tasks a real-world AI assistant or robot would face.