A team led by Tong Zheng and sixteen coauthors has published Dream-RSI, a framework aimed at one of the harder problems in autonomous AI research: teaching agents to explore more efficiently as they hunt for better solutions.
The paper’s core idea is a replay simulator built from an agent’s own history of past discoveries. Rather than trying each new exploration policy in expensive repeated live runs, Dream-RSI scores candidates offline against that record, which the authors call “dreaming.” The refined policy then goes back online, and the loop repeats, expanding the simulator’s pool of discovery history as it improves.
Three domains were tested: GPU kernel work, optimization problems in mathematics, and algorithm engineering. According to the paper, Dream-RSI produced “competitive or improved discovery quality while substantially reducing discovery cost in several settings.” The abstract does not include specific benchmark figures.
The 12-page submission, filed under Computation and Language, adds to a growing body of work on recursive self-improvement, where the bottleneck is increasingly the cost of exploration itself rather than raw compute. Teams building autonomous coding or optimization agents should watch for a fuller results write-up before assuming the cost savings generalize beyond the paper’s three tested domains.
Per the paper’s arXiv abstract, submitted 14 September 2026.
