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

MarkTechPost · 1 month ago · 14

Meta FAIR formalized research preference and introduced AI Research Preference Models (RPMs) to rank unexecuted candidates so an AI research agent executes only the most promising one. The inference-only RPM achieved an average normalized score increase from 0.684 to 0.711, and the agentic RPM reached 0.729 on AIRS-Bench. As a result, the system hit the baseline’s 24-hour results in about 15–15.5 hours and reported new SOTA scores (WinoGrande 94.1% and SVAMP 95.7%).