A research team from FAIR at Meta, the University of Oxford, and University College London has introduced AI Research Preference Models (RPMs) to help AI agents rank unexecuted machine learning experiment candidates. This approach addresses the bottleneck where idea generation is cheap but verification via GPU training is expensive.
- RPMs use frozen pretrained LLMs with no fine-tuning, leveraging the open-source AIRA-dojo scaffold and Qwen3.6-27B as the backbone.
- The system generates 15 parallel candidates and selects one winner through a pairwise knockout tournament grounded in validation scores.
- Two variants were tested: an inference-only judge and an agentic judge that runs short pilot experiments in a sandboxed environment.
- On AIRS-Bench, RPMs improved average normalized scores from 0.684 to 0.711 (inference-only) and 0.729 (agentic).
- Both variants achieved the baseline's final score in roughly 15 hours, providing a 1.5–1.6× speedup over random selection.
The authors consider this important because it allows agents to focus compute resources on the most promising experiments, significantly accelerating research progress without requiring stronger base models.