Google        Google DeepMind        University of Maryland, College Park        University of Virginia

# Dream-RSI: Recursive Self-Improvement
through Evolving Worlds An agent must dream to recursively self-improve. History is the world it dreams in. [![Paper](https://img.shields.io/badge/Paper-PDF-b31b1b.svg)](papers/Dream-RSI.pdf) [![arXiv](https://img.shields.io/badge/arXiv-coming%20soon-b31b1b.svg)](https://www.dream-rsi.com) [![Website](https://img.shields.io/badge/Website-dream--rsi.com-1a73e8.svg)](https://www.dream-rsi.com) [![Tasks](https://img.shields.io/badge/tasks-8%20across%203%20domains-555.svg)](papers/Dream-RSI.pdf) Tong Zheng1,2, Xidong Wu1✉, Zheng Zhang1✉, Zhankui He3, Chaoyi Zhang1, Benjamin Coleman3, Ruoqiao Wei1, Di Bai3, Haolin Liu4, Rui Liu2, Xue Wang1, Yue Zhuan1, Wang-Cheng Kang3, Renkai Xiang1, Heng Huang2, Xinwu Cheng1, Yunsong Guo1 1Google  ·  2University of Maryland, College Park  ·  3Google DeepMind  ·  4University of Virginia ✉ Corresponding authors
--- > [!NOTE] > Code is being prepared for release. See [Release plan](#release-plan).

Algorithm engineering: 1.22x faster downstream runtime, 1.74x less discovery compute, 162x fewer calls than SimpleTES. Mathematical optimization: 2 of 3 tasks at or above the selected baseline. GPU kernel engineering: 4 of 4 kernels improved, 2.09x higher performance at an equal budget, 2.43x fewer generations at equal performance. Zero gradient steps on the coding agent. Versus Recursive Fixed Exploration unless a published system is named; algorithm engineering on Gemini-3.1-Pro.

## News - **Sep 2026** — Preprint and project page released: [paper (PDF)](papers/Dream-RSI.pdf) · [dream-rsi.com](https://www.dream-rsi.com) ## Overview **Progress in recursive self-improvement hinges on effective exploration.** As self-improvement targets get harder, discovery stretches over thousands of proposal–evaluation cycles, and poor exploration wastes substantial computation on ineffective search directions. Managing and improving the exploration strategy is the bottleneck, and current systems face a dilemma: fixed strategies cannot adapt as search spaces scale, while optimizing the policy online means navigating a vast meta-search space under delayed and expensive feedback — assessing one exploration policy requires observing how it shapes an entire discovery process. **Our key insight is that accumulated discovery history can serve as a replay simulator over the realized search space.** A completed discovery process already records a structured tree of past exploration decisions and their realized code-execution outcomes. An alternative policy can traverse that tree differently: different subsets of recorded branches, in different orders, with different parallel groupings and stopping decisions. Because all outcomes are already saved, evaluating it requires only reading past records — no rerunning of the discovery agent or the evaluator. By analogy with model-based RL and world models, the history becomes a world the agent can *dream* in. **Dream-RSI closes a self-improvement loop at the exploration layer.** A lightweight orchestration layer makes exploration explicit and programmable — branching, parallel exploration, stopping — while leaving the underlying coding agent unchanged. ❶ the current policy drives online discovery and logs its traces; ❷ the recorded trees are converted into a reusable simulator pool; ❸ candidate policies are evaluated and refined by dreaming over that pool, which returns immediate, low-cost off-policy feedback. The improved policy is redeployed online, continuously expanding the pool.

The Dream-RSI loop: online explore, construct replay simulator, dreaming-based policy improvement

**[Method, results and interactive walkthrough → dream-rsi.com](https://www.dream-rsi.com)** ## Highlights - **History as replay simulator.** We conceptualize completed discovery histories as replay simulators, which makes delayed exploration feedback reusable for efficient meta-exploration policy evaluation. - **Meta-layer RSI loop.** Dream-RSI 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. - **Empirical validation.** Across algorithm engineering, mathematical optimization and GPU kernel engineering, Dream-RSI improves both discovery effectiveness and efficiency in several settings. ## Release plan | Item | Status | |---|---| | Paper (PDF) | ✅ Available | | Project page & interactive demo | ✅ [dream-rsi.com](https://www.dream-rsi.com) | | arXiv posting | 🔜 In progress | | Discovered programs | ⏳ Being prepared | | Full codebase | ⏳ Being prepared | | Reproduction scripts | ⏳ Being prepared | ## Citation ```bibtex @article{zheng2026dreamrsi, title = {Dream-RSI: Recursive Self-Improvement through Evolving Worlds}, author = {Zheng, Tong and Wu, Xidong and Zhang, Zheng and He, Zhankui and Zhang, Chaoyi and Coleman, Benjamin and Wei, Ruoqiao and Bai, Di and Liu, Haolin and Liu, Rui and Wang, Xue and Zhuan, Yue and Kang, Wang-Cheng and Xiang, Renkai and Huang, Heng and Cheng, Xinwu and Guo, Yunsong}, journal = {arXiv preprint arXiv:2609.14858}, year = {2026} } ```