文章背景与核心概要
在训练长周期终端智能体（Long-horizon terminal agents）时，人工编写训练数据往往面临高昂的成本以及可扩展性受限的瓶颈。为了解决这一痛点，本文作者引入了“递归合成终端任务”（Recursive Synthetic Terminal Tasks, RST）框架。该框架利用递归且经过验证的合成流程，通过扩展种子任务、重新对齐指令与验证器，并在隔离沙箱中进行验证，从而自动生成复杂的终端任务。
在历经 15 轮的递归迭代中，该系统以大约每任务 0.05 美元的极低成本合成了 37,484 个任务。随着轮次的增加，任务的复杂度显著提升，其参考解法和命令数量也大幅增长。在这些合成轨迹上对模型进行微调，使模型在既有基准测试上的性能提升了高达 10 个百分点，这充分证明了递归合成方法为训练先进智能体模型提供了一条可扩展且高效的路径。
Recursive Synthesis for Long-Horizon Terminal Tasks
arXiv: 2608.05466
Authors: Zhongzhi Li, Yucheng Shi, Zongxia Li, Ruhan Wang, Anhao Li, Zixun Huang, Junyao Yang, Lei Ke, Ninghao Liu, Haitao Mi, Leowei Liang
Submitted: 5 Aug 2026 (v1), 7 Aug 2026 (v2)
arXiv: 2608.05466
Authors: Zhongzhi Li, Yucheng Shi, Zongxia Li, Ruhan Wang, Anhao Li, Zixun Huang, Junyao Yang, Lei Ke, Ninghao Liu, Haitao Mi, Leowei Liang
Submitted: 5 Aug 2026 (v1), 7 Aug 2026 (v2)
Summary
The authors introduce Recursive Synthetic Terminal Tasks (RST), a framework designed to overcome the high cost and scalability limitations of human-authored training data for long-horizon terminal agents. By leveraging a recursive, verified synthesis process, the framework generates complex tasks by extending seed tasks, realigning instructions and verifiers, and validating them in isolated sandboxes.
Over 15 recursive rounds, the system generated 37,484 tasks at a cost of approximately $0.05 per task. The complexity of these tasks increased significantly over time, with reference solutions and command counts growing substantially. Fine-tuning models on these trajectories yielded significant performance improvements (up to 10 points) on established benchmarks, demonstrating that the recursive synthesis approach provides a scalable and effective path for training advanced agentic models.
Summary
The authors introduce Recursive Synthetic Terminal Tasks (RST), a framework designed to overcome the high cost and scalability limitations of human-authored training data for long-horizon terminal agents. By leveraging a recursive, verified synthesis process, the framework generates complex tasks by extending seed tasks, realigning instructions and verifiers, and validating them in isolated sandboxes.
Over 15 recursive rounds, the system generated 37,484 tasks at a cost of approximately $0.05 per task. The complexity of these tasks increased significantly over time, with reference solutions and command counts growing substantially. Fine-tuning models on these trajectories yielded significant performance improvements (up to 10 points) on established benchmarks, demonstrating that the recursive synthesis approach provides a scalable and effective path for training advanced agentic models.
Key Contributions
- Scalable Synthesis: RST enables the automated construction of high-quality, long-horizon terminal tasks, bypassing the need for expensive human labor.
- Recursive Methodology: The framework uses accepted tasks from previous rounds as seeds for subsequent rounds, creating a continuous loop of increasing task difficulty.
- Proven Utility: Supervised fine-tuning and agentic PPO on the synthesized data resulted in substantial performance gains for Qwen3.5 models (27B and 122B-A10B) across Terminal-Bench 2, Terminal-Bench Hard, and Long-Horizon Terminal Bench.
- Sustainability: The research indicates no performance ceiling after 15 rounds, suggesting the process can be scaled indefinitely as task difficulty requirements evolve.
Key Contributions
- Scalable Synthesis: RST enables the automated construction of high-quality, long-horizon terminal tasks, bypassing the need for expensive human labor.
- Recursive Methodology: The framework uses accepted tasks from previous rounds as seeds for subsequent rounds, creating a continuous loop of increasing task difficulty.
- Proven Utility: Supervised fine-tuning and agentic PPO on the synthesized data resulted in substantial performance gains for Qwen3.5 models (27B and 122B-A10B) across Terminal-Bench 2, Terminal-Bench Hard, and Long-Horizon Terminal Bench.
- Sustainability: The research indicates no performance ceiling after 15 rounds, suggesting the process can be scaled indefinitely as task difficulty requirements evolve.
Access Paper
Access Paper
License
License