Research

Meta FAIR Introduces AI Research Preference Models to Rank Machine Learning Experiments Before GPU Use

Meta FAIR with Oxford and UCL unveiled RPMs, which rank unexecuted AI experiments and run only the most promising one, cutting GPU hours and improving results.

Story highlights

According to MarkTechPost, Meta's FAIR research team, together with the University of Oxford and University College London, has introduced AI Research Preference Models (RPMs), a method that ranks machine learning experiments before they consume GPU hours. RPMs are designed for AI research agents that can propose, implement and score their own experiments. Generating ideas is cheap, but verifying each one by training can take hours to days of GPU time, so an agent typically has far more candidates than it can afford to run. The RPM formalizes the choice of which candidates to run and selects one for execution, without forecasting an absolute score. The researchers found that language models are unreliable at predicting metrics or execution outcomes, so RPMs rely on relative comparisons.

AI News Feed · 2026-09-07 04:47

The models use frozen pretrained LLMs with no fine-tuning. The scaffold AIRA-dojo and the benchmark AIRS-Bench are open source, and the backbone Qwen3.6-27B has open weights. In the agent loop, the RPM intervenes only at the creation of a new child experiment. Instead of generating one child and executing it, the agent applies an operator 15 times in parallel to create 15 unexecuted candidates, then compares them pairwise in a knockout tournament. Only the winner is executed. Each comparison is grounded in context nodes collected by a breadth-first search walk of the explored tree, with each node showing its validation score.

The system comes in two variants. The inference-only RPM acts as an LLM-as-a-judge, comparing candidate plans, code and search history. Its prompt was optimized with MIPROv2 from DSPy and converged on a principal-investigator rubric that tolerates fixable bugs, rewards extensibility and penalizes redundant directions. The agentic RPM adds a sandbox that clones the agent's environment, including a single H200 GPU, and runs small-scale pilot experiments before selecting the most informative next experiment. Two design choices matter: the remaining budget is deliberately overstated (2,700 seconds reported against a real 300 seconds) to prevent early stopping, and pilots are capped at 30 with a 60-second threshold. Pilot time competes with the agent's own clock, so the agentic selector runs only on Draft and Improve steps; Debug reverts to random selection.

Results on AIRS-Bench, which includes 20 public text and tabular tasks run for 24 hours on a single H200 per task with 10 seeds, showed that the average normalized score was 0.684 with no RPM, 0.711 with inference-only RPM and 0.729 with agentic RPM. A validation oracle reached 0.748 and a test oracle 0.759. The probability of improvement over no RPM was 0.5923 for the inference-only variant and 0.5913 for the agentic variant, with 95% confidence interval lower bounds at 0.5066 and 0.5018. Both variants reached the baseline's final score in about 15 hours, a 1.5–1.6× speedup. Self-hosted inference added 0.660 hours per run, giving an adjusted score of 0.708 at 23.34 hours.

The paper reports two new state-of-the-art results: WinoGrande at 94.1% with the agentic RPM, against a prior agentic SOTA of 90.4%, and SVAMP at 95.7% with inference-only RPM, against a prior human SOTA of 94.2%. The source notes that the gain comes from the selection layer rather than a stronger judge, since the same backbone was used for both operators and the RPM. The project's paper and announcement were shared by MarkTechPost.

Editor's Summary

Meta FAIR and its academic collaborators introduced RPMs, which rank up to 15 unexecuted ML experiment candidates and execute only the most promising one. On AIRS-Bench, both inference-only and agentic RPMs improved average normalized scores and reached a given performance level in roughly 15 hours instead of 24. The method also reported new state-of-the-art results on WinoGrande and SVAMP, with the scaffold, benchmark and model weights open-sourced.