Dream RSI

Dream-RSI (Recursive Self-Improvement through Evolving Worlds) is a 2026 technical report from Google, Google DeepMind and the University of Maryland, College Park, with a few other co-authors. It addresses long-horizon discovery loops, such as coding agents running thousands of proposal–evaluation cycles. In these loops the exploration strategy (where to branch, what to run in parallel, when to stop) is usually hand-written and fixed.

The method treats the recorded discovery tree from a finished run as an exact replay simulator. Each node carries the execution outcome it produced, so a candidate exploration policy can be evaluated by walking the same tree in a different order, with no new executions. Thousands of candidate policies are screened this way offline, and only the best one is deployed online. That deployment records a new tree, which is added to the pool of "worlds", and the cycle repeats. A lightweight orchestration layer makes exploration programmable while leaving the underlying coding agent unchanged.

According to a YouTube creator summary (not verified against the paper), the approach improves the search strategy without changing model weights. It reportedly outperforms other recursive-improvement methods on mathematical optimization and matches GPU kernel results with over 2x fewer generations. The report states that a policy can only be evaluated where history actually went, which is why the process runs as a loop. This is a research project, and the provided material shows no released code, weights or product.