A small AI model can recover 70% of a large model's ability by asking 10 yes-or-no questions

Researchers found that a weak AI model can extract most of what a stronger model knows through an interactive question-and-answer process, using 100 times less data than previous methods. This means smaller models could get smarter by talking to bigger ones instead of needing to process entire responses — useful for phones, robots, or anywhere bandwidth or computing power is tight.

For years, the only way to move knowledge from a large model to a small one was to have the large model generate full text, then compress it. This paper shows that a small model asking binary questions (yes or no, one bit per answer) transfers knowledge far more efficiently. The practical consequence: you could run a capable AI on a phone or embedded device by letting it query a remote model through a thin pipe of binary questions, rather than downloading massive responses or pre-training on huge datasets. This breaks the constraint that capability requires local compute.