The researchers mapped solutions from each run onto a two-dimensional chart where similar ideas sit close together. Solutions from a single run almost always stayed in a small region, which the authors call an idea basin: accumulated chat context can constrain the agent's search, keeping it focused on similar solutions instead of letting it explore. This held even when agents periodically compressed their histories.

The proposed fix, Fork-and-Flush, splits an agent into parallel branches that inherit the workspace, including code, files and saved results, but lose the entire chat context. Branches that kept the parent's chat history tended to drift back to the same idea basin, while fresh context restored diversity in the search. Tested on 13 long-running tasks, from polyomino packing to job-shop scheduling, with runs lasting several hours to over 100 hours, the method scored 0.78 on a min-max-normalised average against 0.47 for single runs and 0.54 for best-of-4, and won the highest mean score on 12 of the 13 tasks.

The study also reports that simply clearing the chat context without branching improved results by 22 percent. Separately, the authors found that variation between runs peaks on tasks at the edge of a model's abilities, and argue that open research problems sit there by definition, so context resets could remain useful even as models improve.