Meta-affiliated team reports model selector cuts AI experiment time

A research team from Meta FAIR, the University of Oxford and University College London reported AI Research Preference Models that rank unexecuted machine-learning experiments before an agent spends compute running them. On its AIRS-Bench evaluation, the team said inference-only and agentic variants reached a no-selector baseline's final score in 14.88 and 15.50 hours, respectively, versus 24 hours. The scaffold and benchmark are open source, while the reported system uses a frozen Qwen3.6-27B backbone. The results are reported by the authors and were measured on 20 public text and tabular tasks using a single H200 per task.