Fork-and-Flush: Escaping Idea Basins in Autoresearch Agents

gist

Microsoft Research authors show that long-running autoresearch agents get stuck: independent runs of the same agent on the same task plateau at very different scores, and an embedding of their artifacts by functional similarity shows each run confined to an “idea basin.” Their fix, fork-and-flush, periodically forks the agent into four branches that keep the workspace but start with a fresh chat context, then continues from the best one. Across 13 tasks lasting up to several days at matched compute, it beats single runs by 66.0% and best-of-4 by 44.4% in normalized average score.

ideas

• Idea basins. Eight runs on one AtCoder task finished between 68.3 and 100.0; an LLM-judged triplet map of their artifacts, matching 85% of held-out judgments, shows each run refining inside one region.
• Fork-and-flush. After an initial four-way start, two fan-outs at 33% and 66% of budget fork four branches from the leader’s files with empty chats, give each about 6% of budget, and keep the winner.
• Both halves matter. Over 11 tasks, flush-only scored 0.55, fork-only 0.57, and the combination 0.71 against 0.45 for a single run; compaction alone did not break the plateaus.
• Visible jumps. On polyomino, a fresh-context branch moved 43 median steps in its first submission into a new basin and finished at 93.9 against a baseline median of 84.6.
• Variance lives at the frontier. A MAX-3-SAT difficulty ladder with Qwen3.5 models shows run-to-run spread peaks at intermediate difficulty, so the authors expect the effect to persist as models improve.

quotes