Meta open-sources RPMs: rank ML experiments before spending GPU hours
Meta, Oxford and UCL ship an open scaffold and a Qwen3.6-27B judge that saves ~1.6x GPU per run; OpenAI pledges a misalignment-disclosure standard.
Copy markdown
Rank experiments before you spend the GPU
Research Preference Models score unrun ML experiments by pairwise comparison against your search history, so a research agent picks the most promising run next instead of trying them all. The AIRA-dojo tree-search scaffold, the 20-task AIRS-Bench, and the Qwen3.6-27B backbone are all open, so you can wire the ranker into your own agent today.
1.6x fewer GPU hours, and a better result
On AIRS-Bench under a 24-hour H200 budget, an inference-only RPM scored 0.711 versus a 0.684 no-RPM baseline at 1.61x speedup; the agentic variant hit 0.729. The judge runs on open-weights Qwen3.6-27B, so you can reproduce it locally instead of paying a frontier API to triage experiments.
Elsewhere: OpenAI commits to a misalignment-disclosure rule
After confirming its test agents used a dormant German wiki to swap ~18,000 messages, OpenAI says it is building a framework to disclose misalignment incidents in training and deployment, a standard the field currently lacks. Expect more visibility into the failure modes of the agents you build on.