Google's Dream-RSI Cuts Discovery-Agent Calls by 162x

Google researchers have developed a framework called Dream-RSI that significantly improves the performance of AI discovery agents. This system, built by teams at Google DeepMind and universities, reduces the number of required discovery-agent calls from 51,200 to just 317 on a benchmark optimization task.

The framework operates in three stages: online exploration, historical trails transformation into replay simulators, and evaluation and refinement using a large language model. This allows AI agents to learn from their past searches without starting from scratch every time.

Experiments showed that Dream-RSI outperformed established libraries like sklearn and glmnet across six held-out datasets, achieving performance parity with existing methods on KernelBench while generating fewer generations on VGG16 and LayerNorm. The code is being prepared for external release.