By replaying the history of its explorations instead of rerunning them, the method reduces the number of calls required by a Scientific Discovery Agent by up to 162 times.
Seventeen researchers from Google, Google DeepMind, the University of Maryland and the University of Virginia describe a method that tackles the cost of automated discovery agents. The principle can be summed up in a formula of the authors: history is the world in which the agent can dream. Rather than restarting a costly exploration with each new strategy tested, the system keeps discovery trees recording every decision made and every result obtained, then simulates candidate strategies on these existing tracks.
The advertised gain is a factor of 162 on the number of agent calls compared to the SimpleTES reference method. On Gemini-3.1-Pro, calls increase from 550 to 317 and the average duration from 3,587 to 2,931 milliseconds. On the ConvDiv task, which focuses on GPU kernel optimization, the number of evaluated attempts gradually decreases from 110 to 50.
The code is published on GitHub. The approach is of interest beyond academic research: automated exploration loops charge for each model call, and a reduction of an order of magnitude changes the economic feasibility of long campaigns.
Read the full article on the original source
Go to the source →