Meta proposes AI Research Preference Models to help agents select high-value experiments

omarsar0 · x · 2026-09-02

Meta discusses a key bottleneck for long-horizon research agents: selecting which experiments to run given limited GPU budget. The paper introduces "AI Research Preference Models" trained to predict the most promising candidate solutions before execution. Two variants are proposed: an inference-only model that reasons over plans and history, and an agentic model that runs small-scale pilot experiments first. Integrated into AIRA-dojo and measured on AIRS-Bench, the approach improved the average normalized score from 0.684 to 0.711 and 0.729, outperforming unguided agents.

Related event: AI Research Preference Models Help Agents Pick Which Experiments Deserve GPU Time(7 posts)→

Original post →

More from coding & agent

coding & agent channel →