RPM Model Teaches AI Agents to Choose the Best Ideas Before...

RPM Model Teaches AI Agents to Choose the Best Ideas Before Expensive Runs

A group of ~33 researchers, including Jason Weston, Anirudh Goyal, and Jakob Foerster, published the paper “AI Research Preference Models” on arXiv about the RPM model, which selects the most promising solution variant before a full run and raises the agent's average normalized score on 20 AIRS-Bench tasks from 0.684 to 0.729.

Compiled by Sergey KostenchukPublished 2026-09-21Updated 2026-09-21

2026-09-21 Research

🤖 RPM Model Teaches AI Agents to Choose the Best Ideas Before Expensive Runs

A team of ~33 researchers, including Jason Weston, Anirudh Goyal, and Jakob Foerster, published the paper “AI Research Preference Models” on arXiv. Their RPM compares multiple solution variants and selects the most promising one before a full run, instead of spending hours of GPU time on each.

🌍 On 20 AIRS-Bench tasks within the AIRA-dojo agent, RPM raised the average normalized score from 0.684 to 0.711 (inference-only) and 0.729 (agentic), and the result of a 24-hour run is achieved in about 15 hours with a budget of less than two-thirds.

👤 The approach can be reproduced in your own pipelines: before an expensive check, add a cheap selection step — reasoning over plans and code, followed by short pilot runs of 5–30 minutes with metric evaluation. The paper is open on arXiv.

Source 1: https://arxiv.org/abs/2608.13940

Source 2: https://arxiv.org/html/2608.13940v2

Sources