Meta's ML Ranking System Cuts GPU Costs by 10%

Training a single machine learning candidate can burn hours or days of GPU time, but generating candidate ideas costs almost nothing. That mismatch means an AI research agent will always propose far more experiments than it can afford to actually run. So the real bottleneck isn't creativity, it's selection: deciding which untested idea deserves the GPU hours before anyone spends them.

A team from Meta FAIR, the University of Oxford and University College London built a system to handle that decision. They call it Research Preference Models, or RPMs, and the framing is deliberate: rather than asking a model to predict a candidate's absolute performance score, which the researchers found language models are bad at doing reliably, RPMs just rank candidates against each other and pick a winner. That ranking happens before execution, inside an evolutionary tree-search scaffold called AIRA-dojo, where the model intervenes right at the point new candidates get created.

The setup is built to be checked and reused rather than taken on faith. AIRA-dojo and the benchmark AIRS-Bench are open source, the Qwen3.6-27B backbone is open weights, and the RPMs themselves run on frozen pretrained LLMs with no fine-tuning involved.