跳转至

文章背景与核心概要

尽管基于终端的代码智能体（Code Agents）正变得日益普及，但现实且可执行的交互式环境依然稀缺，这与海量的智能体轨迹形成了鲜明对比。这种环境的匮乏构成了重大障碍，因为智能体的后训练（Post-training）严重依赖于能够被重新查询、转化为多个可验证任务并提供执行反馈的交互式环境，而单纯的轨迹只是一份被冻结的单一示范。

本文介绍了 Terminal-Universe，这是一个创新的框架，它能够将原始的智能体轨迹转化为可复用、可扩展的终端环境。通过分析工具执行历史，该框架能够重建工作区文件、在广度（跨代码库查询）和深度（多轮交互会话）上合成新任务，并最终产出 37.3k 个任务充分的环境（task-sufficient environments）。在这一语料库上对诸如 Qwen3.5-27B 等模型进行微调，能够在 Terminal-Bench 2.1 和 EvoCode-Bench v2 等基准测试中带来显著的性能提升。

Terminal-Universe：将智能体轨迹转化为可扩展的终端环境 (Terminal-Universe: Turning Agent Trajectories into Scalable Terminal Environments)

arXiv ID: arXiv:2609.04148
Subjects: Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
Submitted: September 3, 2026
Authors: Jie Wu, Zhenru Zhang, Beichen Zhang, Xuwu Wang, Yuhui Su, Mouxiang Chen, Peng Wang, Zhihai Wang, Que Shen, Hao Zhou, An Yang, Fei Huang, Yujiu Yang, Dayiheng Liu

arXiv ID: arXiv:2609.04148
Subjects: Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
Submitted: September 3, 2026
Authors: Jie Wu, Zhenru Zhang, Beichen Zhang, Xuwu Wang, Yuhui Su, Mouxiang Chen, Peng Wang, Zhihai Wang, Que Shen, Hao Zhou, An Yang, Fei Huang, Yujiu Yang, Dayiheng Liu

执行摘要 (Executive Summary)

尽管基于终端的代码智能体正变得日益普及，但现实且可执行的环境依然稀缺，这与海量的智能体轨迹形成了鲜明对比。这种稀缺性带来了重大的障碍，因为智能体的后训练严重依赖于交互式环境——这些环境可以被重新查询为多个可验证的任务并提供执行反馈，而轨迹仅仅是一个单一的、被冻结的示范。

本文介绍了 Terminal-Universe，这是一个新颖的框架，它将原始的智能体轨迹转化为可复用、可扩展的终端环境。通过分析工具执行历史，该框架重建了工作区文件，在广度（跨代码库查询）和深度（多轮交互会话）上合成了新任务，并产出了 37.3k 个任务充分的环境。在这一语料库上对 Qwen3.5-27B 等模型进行微调，在 Terminal-Bench 2.1 和 EvoCode-Bench v2 等基准测试上取得了大幅度的性能提升。

While terminal-based code agents are becoming increasingly prevalent, realistic and executable environments remain scarce despite a surplus of agent trajectories. This scarcity presents a significant hurdle because agent post-training relies heavily on interactive environments that can be re-queried into multiple verifiable tasks and provide execution feedback—whereas a trajectory is merely a single, frozen demonstration.

This paper introduces Terminal-Universe, a novel framework that transforms raw agent trajectories into reusable, scalable terminal environments. By analyzing tool-execution histories, the framework reconstructs workspace files, synthesizes new tasks across breadth (multi-codebase queries) and depth (multi-round interactive sessions), and produces 37.3k task-sufficient environments. Fine-tuning models like Qwen3.5-27B on this corpus yields substantial performance gains on benchmarks such as Terminal-Bench 2.1 and EvoCode-Bench v2.

核心创新与方法论 (Key Innovations & Methodology)

Terminal-Universe 的核心洞察在于：轨迹中记录的工具执行历史本身就暴露了其执行环境的结构与内容。该框架通过几个关键阶段运行：

  1. 工作区重建：
  2. 重放轨迹中记录的文件操作，将每个文件恢复到修改前的状态，从而生成部分工作区。
  3. 采用补全智能体（completion agent）来补充所有缺失的文件和依赖项。
  4. 任务合成：
  5. 重建原始意图任务，同时基于恢复的工作区合成全新的任务。
  6. 双轴扩展：
  7. 广度（Breadth）： 挖掘相关环境之间的定向依赖关系，以合成跨多个代码库的跨工作区查询（模拟现实世界的软件开发）。
  8. 深度（Depth）： 使用用户智能体将单轮查询扩展为多轮会话，以捕捉迭代反馈和需求求精。

The core insight of Terminal-Universe is that a trajectory's recorded tool-execution history inherently exposes the structure and contents of the environment in which it was executed. The framework operates through several key stages:

  1. Workspace Reconstruction:
  2. Replays file operations recorded in a trajectory to restore each file to its pre-modified state, generating a partial workspace.
  3. Employs a completion agent to supply any missing files and dependencies.
  4. Task Synthesis:
  5. Reconstructs original intent tasks while simultaneously synthesizing entirely new ones based on the recovered workspace.
  6. Scaling Along Two Axes:
  7. Breadth: Mines directional dependency relations between related environments to synthesize cross-workspace queries spanning multiple codebases (mirroring real-world software development).
  8. Depth: Extends single-turn queries into multi-round sessions using a user agent to capture iterative feedback and requirement refinement.

结果与性能 (Results & Performance)

  • 环境生成： Terminal-Universe 从公开的终端智能体轨迹中成功生成了 37.3k 个任务充分的环境。
  • 监督微调（SFT）： 在生成的语料库上对 Qwen3.5-27B 进行微调带来了以下成果：
  • 在 Terminal-Bench 2.1 上的单轮性能提升了 11.9 个百分点。
  • 在 EvoCode-Bench v2 MT@4 上的多轮性能提升了 13.8 个百分点。
  • Environment Generation: Terminal-Universe successfully produces 37.3k task-sufficient environments from public terminal agent trajectories.
  • Supervised Fine-Tuning (SFT): Fine-tuning Qwen3.5-27B on the generated corpus results in:
  • An 11.9-point improvement in single-round performance on Terminal-Bench 2.1.
  • A 13.8-point improvement in multi-round performance on EvoCode-Bench v2 MT@4.