Fork-and-Flush: Escaping autoresearch idea basins via population-based ‘training’ 🧑‍🔬 Idea descent: @jackcai1206 @christos_ziakas et al. frame autoresearch as "idea descent" over the space of possible solutions. Like GD, it can get stuck in local optima aka "idea basins". Method: fork the agent into K branches that keep the workspace but start with an empty chat history. Run each for some part of the total budget, keep the best branch, and continue. A single run (never fork) and best-of-N (fork once at the start) are the two extremes of this spectrum. Classical roots: it's basin hopping and population-based training applied to agents. Copy the best solution, then perturb it. The copied state is the workspace, and the perturbation is dropping the chat history. The branch has to work out its intent from the files alone, which is what lets it land in a new basin. Results: at the same compute, fork-and-flush clearly beats both a single long run and best-of-N across a range of research/eng tasks, and results vary less from run to run. Forking and flushing each help on their own, but the combination works best. Interestingly, run-to-run variance (‘forking potential’) is largest when a task is at the edge of model capabilities, so this kind of exploration may stay useful as models get stronger. Paper: arxiv.org/abs/2610.07447