Researchers from Google DeepMind, the University of Maryland, and the University of Virginia introduced Dream-RSI, a framework designed to optimize AI discovery agents. The system enables agents to learn from their own search histories, eliminating the need for frequent, computationally expensive calls to the underlying model.
The framework operates through a three-stage process: generating 'discovery trees' of search paths, converting these into replay simulators, and using a separate large language model to refine exploration policies offline. This approach achieves recursive self-improvement without requiring weight updates or retraining of the core model.
In a Lasso path solver benchmark, Dream-RSI reduced discovery-agent calls from 51,200 to 317, a 162x improvement. The average runtime decreased from 3,587.1 milliseconds to 2,931.0 milliseconds. The system outperformed standard libraries like sklearn and glmnet across six datasets.
Experiments utilized Gemini-3.1-Pro and Gemini-3.7-Flash models without fine-tuning. On KernelBench, the framework achieved performance parity with existing methods while reducing generations by 2.43x on VGG16 and 1.79x on LayerNorm. The research team, led by Tong Zheng, plans to release the code on GitHub.
