Sakhanda Wire
NVDA $153.64 -1.60% MSFT $506.98 -2.13% GOOGL $279.95 -1.33% META $620.09 -1.01% AMZN $256.57 -0.46%
← Back to the news

AI models rank experiments before costly training runs

Researchers from Meta FAIR, the University of Oxford and University College London have introduced AI Research Preference Models (RPMs), designed to choose which machine-learning experiments an AI research agent should execute. The system ranks untested candidates rather than predicting their eventual scores, because the team found language models unreliable at forecasting metrics and execution outcomes.

RPMs generate 15 candidates through the AIRA-dojo search process and compare them in a knockout tournament; only the winner is run. The inference-only version judges plans, code and search history. The agentic version can conduct small pilot experiments in a sandbox with one H200 GPU. Both use frozen pretrained language models without fine-tuning, with Qwen3.6-27B serving as the backbone for the search operators and selector.

On 20 AIRS-Bench text and tabular tasks, the average normalized score rose from 0.684 with random selection to 0.711 for the inference-only RPM and 0.729 for the agentic version. Each reached the random baseline’s 24-hour result in about 15 hours, representing a reported 1.5–1.6× speedup. The team also reported 94.1% on WinoGrande and 95.7% on SVAMP.

Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.

Originally published by MarkTechPost on

Read the original on MarkTechPost ↗

The reporting and the image belong to MarkTechPost and are credited and linked above; the text on this page is Sakhanda Wire's own summary.

← Back to the news

More stories

All the latest news