AI Research Preference Model
A method from Meta FAIR that ranks proposed machine learning experiments against each other, rather than predicting scores, to decide which are worth spending GPU time on.
Research agents can suggest far more experiments than anyone can afford to run, because proposing takes seconds and training takes days. An RPM sits in front of that queue and puts the candidates in order, then picks the top one to actually execute.
The design choice that makes it work is a refusal: it never tries to predict how well an experiment will score, because language models turn out to be unreliable at forecasting absolute numbers. Comparing two options is a much easier question than valuing either one, which is a trick worth borrowing whenever you ask a model to judge something.