AI Research Preference Models Optimize Experiment Budgets.

AI research agents (AIRAs) are becoming capable of proposing, implementing, and evaluating their own machine learning experiments. However, a major bottleneck in advancing frontier tasks is the high cost of evaluation; while candidate solutions can be generated quickly, running them can consume hours or even days of GPU time. This disparity means an agent can propose far more solutions than it can afford to test, making its progress highly dependent on how it prioritizes experiments within a fixed execution budget. To address this, researchers introduce AI Research Preference Models (RPMs). These models are designed to predict which of many candidate solutions are most valuable to execute, without incurring the full computational cost of running them all. RPMs are constructed from frozen pretrained language models and come in two forms: an inference-only model that reasons over plans, code, and prior results, and an agentic model that also conducts small-scale pilot experiments before making a final decision. Integrating these RPMs into the AIRA-dojo search agent and evaluating them on the AIRS-Bench benchmark demonstrated significant improvements. The RPM variants raised the average normalized score and allowed the agent to reach the unguided agent's 24-hour performance in approximately 15 hours, utilizing less than two-thirds of the execution budget. Furthermore, the best RPMs achieved new state-of-the-art results on two AIRS-Bench tasks, highlighting their potential to accelerate AI research.