Meta RPM Model Prioritizes AI Research Experiments, Cuts Compute by 40%

rohanpaul_ai · x · 2026-08-18

Meta FAIR introduces Research Preference Models (RPM) to address the bottleneck where AI agents generate experiment ideas faster than they can afford to run them. AIRA-dojo generates 15 candidates per step; RPM ranks them using code and prior results, sending only the best to the expensive full run. Experiments on AIRS-Bench show average normalized scores rose from 0.684 (random) to 0.711 (inference-only RPM) and 0.729 (agentic RPM), matching baseline 24-hour scores in roughly 15 hours.

Related event: Meta Proposes Research Preference Models to Evaluate AI Experiment Ideas Without Running Them(2 posts)→

Original post →

More from coding & agent

coding & agent channel →