In an 11-task ablation, Fork-and-Flush scored 0.71 on the study’s min–max normalized average measure, ahead of fork-only at 0.57, flush-only at 0.55 and a single run at 0.45. The paper’s authors—Ziyang Cai, Christos Ziakas, Vasilis Kontonis and Tim Pearce—report these results in Fork-and-Flush: Escaping Idea Basins in Autoresearch Agents. The score puts task results on a common scale before averaging.
What Fork-and-Flush clears—and keeps
Autoresearch is an iterative process in which an AI agent proposes a solution, tests it and uses feedback to guide its next attempt. Fork-and-Flush creates parallel branches from the agent’s existing workspace, clears each branch’s inherited conversational context and lets the branches explore for a fixed interval. The workspace and files remain; after scoring, the highest-scoring branch becomes the leader and the others are discarded.
That distinction matters: the reset applies to an agent’s working conversation at a branch point, while project files and other work in the workspace persist. It describes a long-running research-agent workflow, not a setting for deleting a consumer chatbot’s saved conversations.
What the 11-task comparison shows
| Condition | Normalized average score |
| Single run | 0.45 |
| Flush only | 0.55 |
| Fork only | 0.57 |
| Fork-and-Flush | 0.71 |
Both individual components scored above the single-run baseline in this ablation, while the combined method scored highest. The results make the comparison more specific than a claim that clearing context alone accounts for the outcome: the paper tested branching and flushing separately as well as together.
The 13-task benchmark is a separate result
The paper’s main evaluation covered 13 open-ended research and engineering tasks under equal compute budgets. Its min–max normalized average scores were:
| Condition | Normalized average score |
| Fork-and-Flush | 0.78 |
| Single run | 0.47 |
| Best-of-four | 0.54 |
Fork-and-Flush had the highest mean score on 12 of the 13 tasks. Its average normalized standard deviation was 0.16, compared with 0.26 for single-run and 0.21 for best-of-four. The 0.78 result belongs to this 13-task evaluation; the 0.71 score comes from the separate 11-task component ablation.
The tested setup
The experiments used GitHub Copilot CLI version 1.0.73, with the same agent, tools, prompt and task workspace across conditions. The main Fork-and-Flush configuration used four agents per fan-out: an initial round, then additional rounds at 33% and 66% of the total budget. Each branch received about 6% of that budget for exploration.
GPT-5.6-terra, at high reasoning effort, handled 11 tasks. GPT-5.5, at extra-high reasoning effort, handled the kernel and tinyCIFAR tasks. Individual runs lasted up to 107 hours. The suite included problems such as polyomino packing, job-shop scheduling, groundwater-source inference and VLIW-kernel optimization.
Separately, the paper’s capability-boundary analysis tested Qwen3.5 models at 4B, 9B and 27B on random 3-CNF formulas with 10 to 100,000 variables. Version 1 of the paper on arXiv is dated October 5, 2026.
