Meta’s AI Agents Cut Research Time by 40% With New Ranking Model

Generating machine learning ideas is cheap. Verifying them is expensive. A single training run can burn hours of GPU time, creating a bottleneck where agents propose far more candidates than they can afford to test. Researchers from Meta’s FAIR lab, Oxford, and UCL have tackled this inefficiency head-on with Research Preference Models (RPMs). Instead of blindly executing every suggestion, these models rank unexecuted candidates to pick the single most promising one for actual training.

The team integrated RPMs into AIRA-dojo, an evolutionary tree search framework. Here, the agent doesn’t just guess; it generates 15 parallel candidates using operators like Draft or Improve. The RPM then pits them against each other in a knockout tournament, using context from previous experiments to decide the winner. Only that victor gets executed. Crucially, the researchers found that language models are terrible at predicting absolute scores. So, RPMs don’t forecast metrics. They simply compare options pairwise, a method that proved far more reliable for guiding the search.

Two variants emerged from the study. The inference-only RPM acts as a judge, evaluating code and plans without running them. It achieved offline accuracy between 57.7% and 59.0%. The agentic RPM goes further, cloning the environment to run short pilot experiments on a single H200 GPU. To prevent premature stopping, the system deliberately overstates its remaining budget. This aggressive strategy paid off. On the AIRS-Bench benchmark, both RPM types reached the baseline’s 24-hour performance score in roughly 15 hours. That is a 1.5x to 1.6x speedup in research velocity.

The results aren’t just about speed. The agentic RPM hit a new state-of-the-art 94.1% on WinoGrande, beating the prior agentic record of 90.4%. The inference-only model scored 95.7% on SVAMP, surpassing human benchmarks. By using open-weight models like Qwen3.6-27B and avoiding fine-tuning, the team made this approach accessible. It suggests that the next leap in AI research won’t come from bigger models, but from smarter selection. If agents can stop wasting cycles on dead ends, the pace of discovery could accelerate dramatically.