Google DeepMind's Dream-RSI Achieves 162x Reduction in Agent Calls

Google DeepMind unveiled Dream-RSI, an AI recursive self-improvement loop that replays past discoveries to test thousands of exploration strategies cheaply. The method maintains or improves results across algorithm design, math optimization, and GPU kernel engineering, reducing agent calls by up to 162 times.
Key insights
Dream-RSI mechanism and results
Dream-RSI replays past discovery processes, allowing low-cost testing of thousands of exploration strategies before applying the best ones to the next round. In tasks spanning algorithm design, mathematical optimization, and GPU kernel engineering, the approach maintained or improved output quality while reducing the number of agent calls by up to 162 times.