Meta FAIR Introduces AI Research Preference Models (RPMs): Ranking ML Experiments Before Spending GPU Hours

Details

AI research agents can now propose and implement ML experiments autonomously, but GPU time remains the binding constraint. When an agent generates far more experiment candidates than it can train, choosing which ones to run is the real lever on progress. Meta FAIR, working with Oxford and UCL, addresses this directly with Research Preference Models (RPMs) — a selection layer that ranks unexecuted candidates using pairwise comparisons grounded in past experiment results, without ever trying to predict absolute scores. RPMs operate inside AIRA-dojo, an evolutionary tree search framework with Draft, Improve, and Debug operators. Instead of running one generated candidate per step, the system creates 15 parallel candidates and puts them through a knockout tournament; only the winner is trained. Two variants were tested: an inference-only RPM (LLM-as-a-judge with a prompt tuned by MIPROv2 from DSPy, 57.7–59% pairwise accuracy) and an agentic RPM that runs up to 30 small pilot experiments in a sandboxed H200 before deciding. On AIRS-Bench — 20 text and tabular ML tasks, 24 hours per task on a single H200, 10 seeds — random candidate selection scores 0.684. The inference-only RPM reaches 0.711 and the agentic variant hits 0.729, just below the validation oracle ceiling of 0.748. The backbone is Qwen3.6-27B throughout with no fine-tuning required. AIRA-dojo and AIRS-Bench are open source. RPMs drop into an existing agentic ML research pipeline as a plug-in selection layer — the only requirement is the Qwen3.6-27B backbone, which is open weights.