AI Research Preference Models

FAIR at Meta (with Oxford and UCL co-authors) attacks the binding constraint on AI research agents: evaluations that consume days of GPU time, while the agent can propose far more candidates than it can afford to run. Research Preference Models (RPMs) predict which candidate solution is most promising without paying to execute them all. Two variants are built from frozen pretrained LLMs: an inference-only RPM that reasons over candidate plans, code, and previously executed solutions, and an agentic RPM that additionally runs small-scale pilot experiments before judging — mirroring how human researchers cheaply test feasibility before committing to a full run.

Integrated into the AIRA-dojo research agent (with a Qwen3.6-27B backbone for both child creation and selection) and evaluated on the AIRS-Bench machine-learning research benchmark, the two variants lift the average normalized score from 0.684 to 0.711 and 0.729 respectively, reach the unguided agent's 24-hour performance in roughly 15 hours (under two-thirds of the execution budget), and set new state-of-the-art results on two AIRS-Bench tasks. Analysis finds selection accuracy improves when the RPM's justification cites prior evidence, implementation correctness, or the pretrained backbone.