A new AI training technique aims to fix a blind spot in how smaller models learn from bigger ones.
Distillation is how many smaller, cheaper AI models get trained: a 'student' model learns to mimic a larger 'teacher.' A newer variant called on-policy distillation lets the student generate its own responses first, then has the teacher grade and correct them token by token. The problem, according to the paper, is that token-level corrections are fragmented: they nudge individual words back on track but never walk the student through a complete, correct fix. The proposed method, called Step-Level On-Policy Distillation, instead lets the teacher correct longer chunks of the student's own generated path, and the researchers say tuning that chunk length lets it behave like classic supervised fine-tuning at one extreme or like standard on-policy distillation at the other.
Distillation is the standard way most cheaper, faster AI models get made from expensive frontier ones, so a method that fixes a known flaw in that correction process could make training cheaper or more reliable across the board. The paper's own numbers back that up: the new method reportedly beat standard on-policy distillation by 13.4 points in success rate on ALFWorld, a virtual household-task benchmark used to test AI agents.
It's one unreviewed preprint with no independent replication yet, and a household-simulation benchmark is a long way from real-world coding or customer-support agents. Worth watching, not popping champagne over.