Posts describing Dream-RSI say it changes how an agent searches—which branches to explore, when to stop and how to allocate computing resources—rather than changing the underlying model’s weights. According to those accounts, it tests strategies using past exploration records, then deploys a better strategy in the next round. One post claims it matched or improved discovery quality across algorithm design, mathematical optimization and GPU kernel engineering, with agent calls reduced by up to 162× in one setting. Separately, a commenter describes recursive self-improvement as now “just another research topic,” alongside long context and multimodality.