ACuRL accepted to NeurIPS 2026
An ACuRL contributor says the work may have potential for recursive self-improvement, though limited compute has prevented larger-scale experiments.
TLDR
An ACuRL contributor announced the NeurIPS 2026 acceptance on September 25. They reported that improvements held when the policy model itself replaced either the curriculum task generator or the automatic evaluator. They see possible potential for recursive self-improvement, but said limited compute had prevented scaling up those experiments. The announcement also links to the paper and code and says everything has been open-sourced.