Meta FAIR's AI RPMs: Rank ML Experiments, Save GPU Hours

AI Research Preference Models, or RPMs, have been introduced by a research team from FAIR at Meta, the University of Oxford, and University College London. These models rank unexecuted machine learning experiment candidates to decide which ones should run. Here's the thing: training one candidate can take hours or even days of GPU time. So, an AI agent often proposes many more experiments than it can afford to run. RPMs help select the most promising ones. The team found that language models were unreliable at predicting metrics or execution outcomes directly. Instead, an RPM picks a candidate without forecasting an absolute score. What's interesting is that these RPMs use frozen, pretrained large language models with no fine-tuning. Two variants exist: an inference-only RPM and an agentic RPM. The agentic RPM uses a sandbox to run small pilot experiments. Results show that both RPM variants improve performance compared to random selection. The inference-only RPM achieved an average normalized score of 0.711, while the agentic RPM reached 0.729, compared to 0.684 for no RPM. The bottom line is that this approach helps AI research agents make more efficient use of expensive computing resources.