DREAM-RSI: RECURSIVE SELF-IMPROVEMENT THROUGH EVOLVING WORLDS

Summary from the TLDR AI newsletter

Progress in recursive self-improvement hinges on effective exploration. Accumulated discovery history can serve as a replay simulator over a realized search space. Dream-RSI closes a self-improvement loop at the exploration layer. It continuously collects discovery histories through online exploration, constructs replay simulators from them to refine meta-exploration strategies via dreaming, and redeploys the upgraded policy online. Dream-RSI improves both discovery effectiveness and efficiency in several settings.