[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-terminal-universe-trajectory-environments":3,"news-related-c77381d9-29ba-45ff-89df-855d11f90de2":38},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"c77381d9-29ba-45ff-89df-855d11f90de2","Terminal-Universe:Qwen 把旧轨迹反向重建为 3.73 万个环境,27B 微调双基准 +11.9\u002F+13.8 分","终端 agent 的旧轨迹通常用一次就扔。Qwen 团队的 Terminal-Universe 重放轨迹里的文件操作,反向重建出 37.3k 个可执行环境,在其上微调 Qwen3.5-27B,Terminal-Bench 2.1 单轮 +11.9 分,EvoCode-Bench v2 多轮 +13.8 分。","终端代码 agent 越普及,一个资源错配就越显眼:agent 干活的完整记录——轨迹——在批量积累,而 agent 后训练真正稀缺的素材「可执行环境」却始终造不出规模。9 月 3 日提交到 arXiv 的 Terminal-Universe(2609.04148,来自 Qwen 团队等机构的 14 位作者)给出了一个反向思路:环境不必凭空生成,它就藏在旧轨迹里。\n\n论文的判断很直接:环境才是 agent 后训练需要的东西——一个环境可以反复查询、派生出多个可验证任务,还能提供执行反馈;而轨迹只是一次冻结的演示,用一次就废。\n\n## 轨迹是环境的压缩包\n\nTerminal-Universe 的核心观察是:轨迹里记录的工具执行历史,本身就暴露了它当初运行环境的结构与内容。既然如此,环境可以从轨迹反向重建。做法分两步:第一步,重放轨迹中记录的文件操作,把每个文件恢复到 agent 修改之前的状态,得到一个残缺工作区;第二步,派一个补全 agent 补齐缺失的文件和依赖。在复原出的工作区上,既能重建原始意图任务,也能合成全新任务。\n\n## 广度与深度两条扩展轴\n\n光有环境还不够,论文沿两个方向扩任务。广度上,在相关环境之间挖掘有向依赖关系,合成跨多个代码库的 cross-workspace 查询——对应真实开发里「改这个库还得看那个库」的常态。深度上,把单轮查询扩成多轮会话,由一个用户 agent 模拟迭代的用户反馈与需求变更。\n\n## 数字盘点\n\n应用到公开终端 agent 轨迹上,该框架产出 37.3k 个「任务充分」的环境。在这批语料上对 Qwen3.5-27B 做监督微调,Terminal-Bench 2.1 单轮成绩提升 11.9 分,EvoCode-Bench v2 MT@4 多轮成绩提升 13.8 分。论文发布次日即登上 Hugging Face Daily Papers 当日第一,拿到 182 个 upvote。\n\n## 值得注意的三点\n\n一是轨迹可能是 agent 时代最被浪费的资产。各家团队都在攒轨迹,SFT 用完一次就扔;这篇工作证明轨迹是环境的「压缩包」,能反向解出可复用、可再出题的工作区。二是真实性优势:文件结构来自真实运行过的环境,不是模型凭空想象,比从零生成环境的路线少了一层幻觉风险。三是接口意义:可重复查询加执行反馈,恰好是把训练从 SFT 推向 RL 所需的闭环条件——论文本身只做了 SFT,但地基明显指向更远处。\n\n对正在积累海量 agent 日志的团队来说,这篇论文的启示很实际:你的轨迹库里可能已经躺着一批免费的环境,只是还没人去解压。\n\n原文:https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.04148\n","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.04148","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"c187600e-804c-4697-b828-1e4330e0eb10","qwen",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"2e147e83-b47b-440c-bf1b-c85f7fbd384a","en","Qwen Turns Old Agent Trajectories Into 37.3k Training Environments","Qwen's Terminal-Universe turns agent trajectories into 37.3k environments; SFT lifts Qwen3.5-27B +11.9 on Terminal-Bench 2.1 and +13.8 on EvoCode-Bench v2.","As terminal-based code agents spread, an odd resource mismatch is widening: agent trajectories — full records of agent work sessions — are accumulating at scale, while the thing agent post-training actually needs, realistic executable environments, remains scarce. Terminal-Universe (arXiv:2609.04148, submitted Sep 3 by 14 authors from the Qwen team and partner institutions) takes the reverse route: environments do not have to be generated from scratch, because they are already hiding inside old trajectories.\n\nThe paper's premise is blunt. An environment can be re-queried into many verifiable tasks and provides execution feedback, whereas a trajectory is a single frozen demonstration — used once, then spent.\n\n## Trajectories Are Compressed Environments\n\nThe core observation: the tool-execution history recorded in a trajectory already exposes the structure and contents of the environment it ran in. So environments can be reconstructed backwards. Step one replays the file operations recorded in the trajectory, restoring each file to its pre-modification state and yielding a partial workspace. Step two dispatches a completion agent to supply the missing files and dependencies. On the recovered workspace, the framework both reconstructs the original intent task and synthesizes entirely new ones.\n\n## Two Scaling Axes: Breadth and Depth\n\nEnvironments alone are not enough, so tasks are scaled along two axes. For breadth, the method mines directional dependency relations between related environments and synthesizes cross-workspace queries spanning multiple codebases — mirroring how real developers routinely touch several repos in one change. For depth, the initial single-turn query is extended into a multi-round session where a user agent simulates iterative feedback and requirement refinement.\n\n## The Numbers\n\nApplied to public terminal agent trajectories, the framework produces 37.3k task-sufficient environments. Supervised fine-tuning of Qwen3.5-27B on this corpus improves single-round performance on Terminal-Bench 2.1 by 11.9 points and multi-round performance on EvoCode-Bench v2 MT@4 by 13.8 points. The day after release it ranked #1 on Hugging Face Daily Papers with 182 upvotes.\n\n## Three Takeaways\n\nFirst, trajectories may be the most wasted asset of the agent era. Teams hoard them, burn them once in SFT, and move on; this work shows a trajectory is a compressed environment waiting to be unpacked into reusable, re-queryable workspaces. Second, authenticity: file structures come from environments that actually ran, not from model imagination — one less layer of hallucination risk compared to generating environments from scratch. Third, the interface matters: re-queryable tasks plus execution feedback are exactly the closed-loop ingredients for pushing post-training from SFT toward RL. The paper only runs SFT, but the foundation clearly points further.\n\nFor any team sitting on mountains of agent logs, the practical hint is simple: your trajectory store may already contain a batch of free environments — nobody has unpacked them yet.\n\nPaper: https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.04148\n","terminal-universe-trajectory-environments","2026-09-04T17:10:00Z","2026-09-04T17:07:24.769225Z","2026-09-04T17:07:24.769233Z",true,"agent",37,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"63c30bcd-3ffc-47c5-bd74-c2a9ed8f7c94","DeepSeek Harness 预览版开源:Agent 被拆成可插拔的插件栈,模型只负责想、Harness 负责做事","deepseek-harness-plugin-stack","2026-09-05T06:00:00+00:00",{"id":46,"title":47,"news_slug":48,"published_at":49},"14a7f5ab-e270-461c-b862-4bde139e463f","HarnessDev 基准:让 LLM 自建 Agent Harness,代码领域仍输人类工程师","harnessdev-llm-selfbuilt-agent-harness","2026-09-03T19:10:00+00:00",{"id":51,"title":52,"news_slug":53,"published_at":54},"6b203495-fcab-4afe-baa7-1079cf993796","拆开 GLM-5.3 的「后训练工厂」:基座一字未动,靠环境合成与 1e-7 对齐撑起全部提升","glm-5-3-post-training-stack-deep-dive","2026-08-17T13:00:00+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"edefe1ba-ee28-4f3a-94f4-ab898e079807","ZCode 提示词泄露:39 万字符暴露 GLM-5.3 智能体的 Claude Code 血统","zcode-391k-prompt-leak-claude-code-dna","2026-08-16T17:15:00+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"abccd085-d910-4846-aae1-5f2136238f3b","别怪大模型:Coding Agent 真正的瓶颈在「脚手架」——Queen's 35 版本控制变量实证","coding-agent-scaffold-bottleneck","2026-07-04T03:55:25+00:00",{"id":66,"title":67,"news_slug":68,"published_at":69},"4860cd5e-0e85-458a-bb63-cf87e137332b","MCP 2026-07-28 Release Candidate：协议层正式无状态化，Agent 后端跑上普通 HTTP 基础设施","mcp-2026-07-28-stateless","2026-07-01T08:20:00+00:00"]