Microsoft's Fork-and-Flush lifts agent scores to 0.78

The study looked at what Microsoft researchers call 'Autoresearch agents' — systems that repeatedly propose solutions, evaluate them, and refine them based on feedback. Eight independent runs of the same agent on the same task ended at markedly different performance levels, and the researchers plotted the solutions on a two-dimensional map that places similar solutions close together. Solutions from a single run almost always stayed within a small region, which the authors call an 'idea basin'.

The proposed fix combines two mechanisms. In forking, the agent splits into K parallel branches, each inheriting the previous workspace and all generated files. In flushing, each branch loses its entire chat context but keeps the workspace, so it can start a fresh conversation and build on prior work. A branch that keeps the parent process's chat context tends to return to the same idea basin because the previous reasoning steers its next steps, whereas fresh context restores diversity in the search.

The researchers also tested whether more capable models avoid idea basins, using MAX-3-SAT problems across three model sizes. Variance between runs followed an inverted U-shaped curve, peaking on tasks at the edge of a model's abilities. The authors argue that open research problems sit, by definition, at the limits of models' abilities, so the variation Fork-and-Flush exploits may persist as models improve, with the boundary simply shifting toward harder problems. Simply clearing the chat context without branching improved results by 22 percent.