Dream-RSI

A Google-led framework for the exploration bottleneck in recursive self-improvement (ten of its 17 authors, including the last author, are at Google; the first author is at Google and the University of Maryland, with Google DeepMind and UVA co-authors): fixed exploration strategies fail as search spaces grow, while optimising them online means searching a vast meta-space under delayed, expensive feedback. Dream-RSI makes exploration explicit and programmable through a lightweight orchestration layer and trains exploration policies inside replayable "dream" simulators of the task world rather than in the expensive world itself, leaving the underlying agent untouched. Evaluated on algorithm engineering, mathematical optimization and GPU kernel engineering. The repo passed 1,300 GitHub stars within three weeks.