AI agents that design their own machine learning experiments have a spending problem: writing a candidate experiment takes minutes, but running it to see if it works can eat hours or days of GPU time. A new paper proposes giving these agents better judgment about which candidates are worth that cost.
Researchers built what they call AI Research Preference Models, or RPMs, which predict how promising a candidate experiment is before anyone spends compute running it. Two versions were tested: one that just reasons over the proposed code and past results, and a more expensive one that also runs small pilot tests first. Both were plugged into an existing research agent called AIRA-dojo and evaluated on AIRS-Bench, a benchmark of ML research tasks. The reasoning-only version pushed the benchmark's normalized score from 0.684 to 0.711; the version that runs pilots first did better, hitting 0.729 and setting new state-of-the-art marks on two of the benchmark's tasks.
The more interesting number is efficiency, not just accuracy. With an RPM guiding it, the agent matched the unguided version's 24-hour results in about 15 hours, using less than two-thirds of the original compute budget - a cut of over a third in GPU time for the same outcome. For labs running these agents at scale, that is the difference between a research budget that stretches to more ideas and one that doesn't.
It's a small, unglamorous fix, but the field has plenty of those: as agents get better at writing candidate solutions, the bottleneck quietly shifts to deciding which ones deserve a GPU. Prioritization, not generation, may end up being the harder problem.