A Microsoft Research paper posted to arXiv on October 5, “Fork-and-Flush: Escaping Idea Basins in Autoresearch Agents,” reports that autoresearch agents can get stuck in “idea basins”: running the same agent on the same task eight times yielded final scores from 68.3 to 100.0 (median 81.9), and throwing extra compute at the weak runs did not close the gap. The proposed Fork-and-Flush method copies the agent into parallel branches that inherit the full workspace — code, files and experiment results — but start with a cleared chat history, then keeps the best-scoring branch after a fixed period. Across 13 long-horizon tasks under the same compute budget, the normalized average score rose to 0.78 versus 0.47 for a single run and 0.54 for best-of-four, with standard deviation falling from 0.26 to 0.16; in ablations, flushing alone gave roughly a 22% gain while forking and flushing together worked best. Tests with Qwen3.5 at 4B, 9B and 27B on MAX-3-SAT showed run-to-run variance peaks near the model’s capability edge, at about 0.5 attainment, suggesting the effect may persist as models improve.

Microsoft Research·arXiv·GitHub Copilot·+3