{"id":116203,"date":"2026-09-07T01:25:26","date_gmt":"2026-09-07T01:25:26","guid":{"rendered":"https:\/\/youzum.net\/meta-fair-introduces-ai-research-preference-models-rpms-ranking-ml-experiments-before-spending-gpu-hours\/"},"modified":"2026-09-07T01:25:26","modified_gmt":"2026-09-07T01:25:26","slug":"meta-fair-introduces-ai-research-preference-models-rpms-ranking-ml-experiments-before-spending-gpu-hours","status":"publish","type":"post","link":"https:\/\/youzum.net\/fr\/meta-fair-introduces-ai-research-preference-models-rpms-ranking-ml-experiments-before-spending-gpu-hours\/","title":{"rendered":"Meta FAIR Introduces AI Research Preference Models (RPMs): Ranking ML Experiments Before Spending GPU Hours"},"content":{"rendered":"

AI research agents can already propose, implement and score their own machine learning experiments. Idea generation is cheap; verification is not. Training one candidate can consume hours to days of GPU time, so an agent proposes far more candidates than it can afford to run. Which ones get run is the real lever on research progress.<\/p>\n

A research team from FAIR at Meta, the University of Oxford and University College London formalizes that lever as research preference<\/strong> and introduces AI Research Preference Models (RPMs)<\/strong>. An RPM ranks unexecuted candidates and picks one to execute. It never forecasts an absolute score, the team found language models unreliable at predicting metrics or execution outcomes.<\/p>\n

Is it deployable?<\/strong> Partially. RPMs use frozen pretrained LLMs with no fine-tuning, the scaffold AIRA-dojo<\/a> and benchmark AIRS-Bench<\/a> are open source, and the backbone Qwen3.6-27B<\/a> is open weights. <\/p>\n

Where the RPM sits in the agent loop<\/strong><\/h2>\n

AIRA-dojo<\/a> is an evolutionary tree search: greedy parent selection, Draft \/ Improve \/ Debug operators, highest-validation-score node returned at the end. The RPM intervenes at child creation<\/strong> only. Instead of generating one child and executing it, the agent applies the operator 15 times in parallel to yield 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 BFS walk of the explored tree, each shown with the validation score it obtained.<\/p>\n

\n<\/div>\n

\n

Two variants, two compute budgets<\/strong><\/h2>\n<\/p>

Inference-only RPM<\/strong>: An LLM-as-a-judge over candidate plans, code and search history. Its prompt was optimized with MIPROv2 from DSPy<\/a>, converging on a principal-investigator rubric that tolerates fixable bugs, rewards extensibility and penalizes redundant directions, offline accuracy 57.7% to 59.0%.<\/p>\n

Agentic RPM<\/strong>: The same judge and a sandbox that clones the agent\u2019s environment, including a single H200. Tools are python<\/code>, bash<\/code> and submit_solution<\/code>. It runs small-scale pilot experiments, then a feedback model either proposes the most informative next experiment or ends the loop. Two design choices carry weight: the remaining budget is deliberately overstated (2,700s reported against a real 300s) so the agent does not stop early, and pilots are capped at 30 with a 60-second threshold. Pilot time competes with the agent\u2019s own clock, so the agentic selector runs only on Draft and Improve steps; Debug reverts to random.<\/p>\n

Results on AIRS-Bench<\/strong><\/h2>\n

Setup: 20 public text and tabular tasks, 24 hours on a single H200 per task, 10 seeds, Qwen3.6-27B as backbone for both the operators and the RPM, so the gain comes from the selection layer, not a stronger judge.<\/p>\n

\n\n\n\n\n\n\n\n\n\n
Child selection<\/th>\nAvg. normalized score<\/th>\n<\/tr>\n<\/thead>\n
No RPM (random pick)<\/td>\n0.684<\/td>\n<\/tr>\n
Inference-only RPM<\/td>\n0.711<\/td>\n<\/tr>\n
Agentic RPM<\/td>\n0.729<\/td>\n<\/tr>\n
Validation oracle (ceiling)<\/td>\n0.748<\/td>\n<\/tr>\n
Test oracle (ceiling)<\/td>\n0.759<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/figure>\n

Probability of improvement over No-RPM is 0.5923 and 0.5913, with 95% CI lower bounds at 0.5066 and 0.5018.<\/p>\n

Efficiency is the more practical result. Inference-only reaches the baseline\u2019s final 0.684 in 14.88 hours (1.61\u00d7), agentic in 15.50 hours (1.55\u00d7). Self-hosted inference adds 0.660 hours per run; adjusting for it still gives 0.708 at 23.34 hours.<\/p>\n

Two new reported SOTA results: WinoGrande 94.1%<\/strong> with the Agentic RPM against a prior agentic SOTA of 90.4% from AIRA\u2082<\/a>, and SVAMP 95.7%<\/strong> with inference-only against a prior human SOTA of 94.2%.<\/p>\n

Key Takeaways<\/strong><\/h2>\n