Clearing chat history helps AI research agents find better solutions
In these tests, resetting chat context helped research agents explore more solutions and achieve higher, more consistent scores on the same compute budget.
• In a Microsoft Research study, Fork-and-Flush scored 0.78 across 13 tasks, versus 0.47 for single runs on the same compute budget.
• The method starts parallel branches with fresh chat contexts while keeping their workspace files and results.
• Across three Qwen3.5 model sizes, run-to-run variation peaked on MAX-3-SAT tasks near each model’s ability limit.
Read more
October 13, 2026 25
