Microsoft Study Finds Fork-and-Flush Improves Long-Running AI Agents’ Solutions

Microsoft Research reported that its Fork-and-Flush method helped long-running AI research agents find better solutions by creating parallel branches that kept workspace files but discarded chat history. Across 13 tasks under the same compute budget, the method achieved a min-max-normalized average score of 0.78, compared with 0.47 for single runs and 0.54 for best-of-4. The study examined agents that repeatedly propose solutions, evaluate them and refine them using feedback. The researchers found that accumulated conversation context can keep agents returning to similar approaches, even when their histories are periodically compressed. Fork-and-Flush lets branches retain work saved in files while starting with a cleared conversation history.

Generated by AI from reports · Updated 5 days ago