Google's Dream-RSI cuts Gemini calls 42% on a Lasso task by evolving only its search policy

Each search runs through one method. In the minimal structure printed in Appendix B.2, `solve()` resets the question and then loops: read what has been revealed so far, update the set of closed branches, select a batch, probe it, record the result curve ^[7]. An empty batch breaks the loop. The loop condition is `while not _budget_done(question, budget):`, so the policy runs until the budget is spent ^[7].

The system rewrites only that policy ^[4]. Gemini, working through Gemini CLI, writes candidate code. A fixed evaluator scores it, and a second frozen LLM rewrites the policy ^[5]. In the abstract, the authors wrote: "A lightweight orchestration layer makes exploration explicit and programmable while leaving the underlying coding agent unchanged." ^[12]

Google DeepMind announced AlphaEvolve in May 2025. It already proposed programs, scored them, kept the best and mutated them, with the model held fixed ^[14]. Dream-RSI adds one layer on top of that kind of loop ^[14].

The headline cut checks out: 233 fewer calls out of 550 is 42.4% ^[1]. The runtime gain on the same task works out to about 18% ^[2]. On Lasso, then, the policy got a better program out of less search. The circle-packing tie with AlphaEvolve V2 is harder to price, because the dev.to write-up does not give a call count for that run ^[10].

Eliezer Yudkowsky fixed the safety community's meaning of the term in a 2008 LessWrong essay. In that sense a seed AI rewrites its own source or weights, and the thing that improves is the thing doing the improving ^[13]. Here a different model rewrites the policy, and that model never changes. According to the dev.to write-up, the model's weights never move anywhere in the 36-page paper ^[3][1]. The X posts saying Google had "cracked" recursive self-improvement ^[2] appear to have stopped at the title. The write-up's own verdict: "A discount; nothing here looks like takeoff." ^[16]

I think that verdict is right, and the discount is still worth having. The 42% is a claim about one Lasso task scored by a fixed evaluator ^[8][5]. Three things have to hold before it transfers to another team's loop. The scorer has to be cheap enough to run on every candidate. Coder calls have to dominate the bill. And the search tree needs branches worth closing early.

The design choice worth copying is making exploration a program. A Python policy can be read and diffed. Because the coding agent is left unchanged ^[12], the policy can sit on an existing loop without retraining anything. The dev.to write-up suggests asking three things of any "RSI" paper: what object changes, who is frozen, and what the headline number is a unit of ^[15]. For Dream-RSI the answers are a Python search policy, all three models, and Gemini calls ^[4][5][8].