AI research agents can already propose, implement, and score their own machine learning experiments. However, idea generation is cheap; verification is not. Training a single candidate can consume hours to days of GPU time, so an agent typically proposes far more candidates than it can afford to run. The selection of which candidates to execute is the real lever on research progress.
A research team from Meta FAIR, the University of Oxford, and University College London formalizes this selection as research preference and introduces AI Research Preference Models (RPMs). An RPM ranks unexecuted candidates and selects one to execute. It never forecasts an absolute score—the team found that language models are unreliable at predicting metrics or execution outcomes.
Is it deployable? Partially. RPMs use frozen pretrained LLMs with no fine-tuning. The scaffold AIRA-dojo and benchmark AIRS-Bench are open source, and the backbone Qwen3.6-27B has open weights.
Where the RPM Sits in the Agent Loop
AIRA-dojo is an evolutionary tree search: it uses greedy parent selection, Draft / Improve / Debug operators, and returns the node with the highest validation score at the end. The RPM intervenes at child creation only. Instead of generating a single child and executing it, the agent applies the operator 15 times in parallel, producing 15 candidate children. The RPM then ranks these candidates and selects the most promising one to actually execute, saving GPU hours that would otherwise be spent evaluating less promising paths.
This approach, detailed in a paper and blog post, aligns with broader 2026 trends in AI research efficiency, where foundation models are increasingly used as decision-makers to cut down on expensive training runs.
via MarkTechPost