Google Deepmind's Dream-RSI helps AI agents improve by “dreaming” about past attempts
Google DeepMind's Dream-RSI improves AI agent search strategies by replaying recorded search histories, cutting attempts and compute across code, math, and GPU kernel tasks.
Google DeepMind researchers introduced Dream-RSI, a method that optimizes agent search strategies by replaying recorded search histories against thousands of alternative decisions without re-calling the model or evaluator. Tested with Gemini 3.1 Pro and Gemini 3.7 Flash on eight tasks, it cut attempts on a genomics/finance optimization from 550 to 317 and beat sklearn and glmnet on all six test datasets, outperforming SimpleTES, which needed 51,200 runs. On GPU kernel tasks it matched performance with up to 2.43x fewer generations or achieved up to 2.09x higher performance within the same budget. The work builds on DeepMind's AlphaEvolve and related efforts like WikiSkill amid growing attention to recursive self-improvement.
- Replays stored search trees to test thousands of alternative strategies without new model calls
- Outperformed SimpleTES using 317 attempts versus 51,200 runs on one task
- GPU kernel tasks: up to 2.43x fewer generations or 2.09x higher performance at equal budget
- Condensing histories into explicit instructions hurt performance by over-narrowing exploration
- Code released on GitHub; operates one level above AlphaEvolve by optimizing search strategy itself
- Tags
- ai-agentscode-generationdream-rsigeminigoogle-deepmindgpu-kernelsrecursive-self-improvementsearch-strategy
- Vendors
- GoogleGoogle DeepMind
- Organizations
- Google DeepMindAnthropic
- AI models
- Gemini 3.1 ProGemini 3.7 Flash
From the original article
4 min read at the-decoder.com
Google and Deepmind's Dream-RSI lets AI agents "dream" through past search runs to test new strategies without costly recalculations. In tests, it matched or beat existing results, cutting iterations by a factor of up to 2. 43.
By Jonathan Kemper · the-decoder.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.