TMax:AllenAI 最强开源终端 Agent RL 配方,9B 参数 Terminal-Bench 2.0 达 27%
TMax: AllenAI's Strongest Open RL Recipe for Terminal Agents — 9B Hits 27% on Terminal-Bench 2.0
by Mycelium Protocol
终端 Agent(Terminal-using agents)——让语言模型直接操作 Shell、跑命令、完成真实计算机任务——已经成为当前最热门的 LM 下游应用。但与这个方向的热度形成反差的是:基于 RL 的训练方法在学术上几乎是空白。
难 benchmark、缺数据、没有可复现的基础配方,是三道门槛。
AllenAI 的 TMax(Hamish Ivison 等)正面拆掉这三道门槛:开源完整数据集、训练代码、模型权重,并给出一个用 9B 参数在 Terminal-Bench 2.0 达到 27% 的配方——超越此前更大规模的闭源模型。
核心结果
TMax-9B(Qwen3.5-9B,RL 微调)在 Terminal-Bench 2.0 达到 27%,这个数字的含义:
- 只用 9B 参数,超过先前更大规模的工作
- Terminal-Bench 2.0 是当前公认最严格的终端 Agent 基准之一(真实环境、程序化验证器)
- 配方简单:outcome-only RL(结果信号,不用 process reward),没有花哨的中间步骤奖励
四阶段配方
整个 TMax 系统围绕四个模块:
1. 数据生成(rl_data/)
数据是最关键的贡献之一。TMax 设计了一个组合采样器,把终端任务表示为若干正交维度的笛卡尔积:
- 难度控制:在生成时显式标注任务难度,让模型训练样本的难度分布可调
- 人设多样化(Personas):让同一类任务在不同使用场景下变形,增加泛化性
- 验证器多样化:程序化验证器覆盖更广的任务类型,避免单一验证模式的过拟合
四阶段流水线:
生成任务 → 用 LLM Agent pass@k 求解 → 分析通过率和语料平衡性 → 上传到 HuggingFace Hub
每个任务都被打包成自包含的 Apptainer/Docker 环境,附带程序化验证器——可以直接拿去让任何 Agent 跑评测,不需要重新搭环境。
最终语料:15K 任务(10K 传统任务 + 5K 多模态复杂任务),比此前最大的终端 Agent 数据集大 2.5 倍。
2. Agent(Vanillux2Agent/)
训练和评测用的 Agent 叫 Vanillux2Agent:
- 基于 LiteLLM,直接调用语言模型
- Prompt 框架来自 mini-SWE-agent(bash 工具 schema、submit 标记、格式错误恢复、输出截断)
- 通过 Harbor 的沙箱环境执行命令
设计哲学:足够简单。没有多余的 Orchestration 层,让模型能力而不是 Agent 框架复杂度决定最终表现。
3. 训练(training/open-instruct/)
基于 open-instruct 的 fork,修复了 Qwen3.5 的若干问题:
| 阶段 | 方法 | 说明 |
|---|---|---|
| SFT warm-start | Supervised Fine-Tuning | 用 pass@k 求解成功的轨迹做冷启动 |
| RL 训练 | DPPO(Distributed PPO) | outcome-only 奖励,不依赖中间过程信号 |
训练的模型系列:
- TMax-4B:Qwen3.5-4B + RL
- TMax-9B:Qwen3.5-9B + RL(主力模型,27% on TB 2.0)
4. 评测(scripts/ + beaker_configs/)
评测基准:
- Terminal-Bench 2.0(主要基准)
- TB-Lite(轻量版)
- SWE-bench(代码修复)
本地复现最简路径:
# 用 vLLM 本地起服务
uvx vllm==0.19.1 serve allenai/tmax-9b \
--served-model-name tmax-9b \
--enable-auto-tool-choice \
--tool-call-parser qwen3_xml \
--tensor-parallel-size 8 --port 8008
# 用 Harbor 跑 Terminal-Bench
uv run harbor run \
--dataset [email protected] \
--env daytona \
--agent-import-path Vanillux2Agent:Vanillux2Agent \
--model openai/tmax-9b \
--agent-kwarg api_base=http://localhost:8008/v1 \
-k 5
开源内容
| 内容 | 位置 |
|---|---|
| 代码(数据生成 + 训练 + 评测) | github.com/hamishivi/tmax |
| 模型(TMax-4B、TMax-9B) | HF: allenai/tmax |
| 数据集(TMax-15K-Harbor) | Harbor 注册表 |
| 论文 | arXiv 2606.23321 |
许可证:Apache 2.0。
为什么值得关注
终端 Agent 的 RL 训练此前几乎没有可复现的开源基础。TMax 填补了这个空缺,而且不是通过堆规模:9B 参数、简单的 outcome-only RL、一套可扩展的数据生成流程。
数据生成部分的设计——“把任务看作正交维度的笛卡尔积”——是一个值得借鉴的思路。它让语料的难度和类型分布变成了可控参数,而不是靠抓取人类数据碰运气。
对于想在终端 Agent 方向做研究的团队,TMax 是目前最好的起点:有基准数据、有可扩展的数据生成、有可复现的训练配方、有开源模型权重做 baseline 对比。
仓库:github.com/hamishivi/tmax · 论文:arXiv 2606.23321 · 模型:allenai/tmax
Mycelium Protocol — 追踪 AI 系统的底层演化
关于 Mycelium
菌丝协议。持续追踪 AI 工具、系统和实验的内容节点。
TMax: AllenAI’s Open RL Recipe for Terminal Agents — 27% on Terminal-Bench 2.0 with 9B Parameters
by Mycelium Protocol
Terminal-using agents — language models that operate a shell, run commands, and complete real computer tasks — have become the most popular downstream application of LMs. Yet despite this popularity, RL-based training for terminal agents remains largely uncharted academic territory. The barriers: difficult benchmarks, scarce data, and no reproducible baseline recipe.
TMax (AllenAI, Hamish Ivison et al.) dismantles all three: open-source dataset, training code, and model weights — plus a recipe that achieves 27% on Terminal-Bench 2.0 with only 9B parameters, outperforming larger models from prior work.
Core Results
TMax-9B (Qwen3.5-9B, RL fine-tuned) achieves 27% on Terminal-Bench 2.0:
- Only 9B parameters, outperforming larger prior models
- Terminal-Bench 2.0 is the field’s most rigorous benchmark (real environments, programmatic verifiers)
- Outcome-only RL — no process reward, no intermediate step supervision; just final task success
The Four-Stage Recipe
1. Data Generation (rl_data/)
The data contribution may be the most important. TMax represents terminal tasks as a Cartesian product of orthogonal axes, making the corpus composition a controllable variable rather than a fixed artifact of data collection:
- Difficulty control: explicit difficulty tagging at generation time
- Persona diversification: the same task class varies across usage contexts for better generalization
- Verifier diversification: programmatic verifiers across diverse task types prevent overfitting to any single verification pattern
Four-stage pipeline:
generate_tasks → solve at pass@k → analyze → upload to HuggingFace Hub
Every task ships as a self-contained Apptainer/Docker environment with a programmatic verifier — usable out of the box, no rebuild required.
Final corpus: 15K tasks (10K traditional + 5K intricate multi-modal), 2.5× larger than the largest prior terminal-agent dataset.
2. Agent (Vanillux2Agent/)
The training and evaluation agent, Vanillux2Agent, is deliberately minimal:
- LiteLLM-based, calls the language model directly
- Prompt framework from mini-SWE-agent (bash tool schema, submit marker, format-error recovery, output truncation)
- Executes commands through Harbor sandboxes
Philosophy: keep the agent simple so that model capability — not agent orchestration complexity — determines performance.
3. Training (training/open-instruct/)
Fork of open-instruct with Qwen3.5 fixes:
| Stage | Method | Notes |
|---|---|---|
| SFT warm-start | Supervised Fine-Tuning | Successful pass@k trajectories as cold-start data |
| RL | DPPO (Distributed PPO) | Outcome-only reward signal |
Models released:
- TMax-4B: Qwen3.5-4B + RL
- TMax-9B: Qwen3.5-9B + RL (main model, 27% on TB 2.0)
4. Evaluation
Benchmarks: Terminal-Bench 2.0, TB-Lite, SWE-bench.
Quick local eval:
# Serve with vLLM
uvx vllm==0.19.1 serve allenai/tmax-9b \
--enable-auto-tool-choice --tool-call-parser qwen3_xml \
--tensor-parallel-size 8 --port 8008
# Run Terminal-Bench via Harbor
uv run harbor run \
--dataset [email protected] \
--agent-import-path Vanillux2Agent:Vanillux2Agent \
--model openai/tmax-9b \
--agent-kwarg api_base=http://localhost:8008/v1 \
-k 5
Open-Source Checklist
| Artifact | Location |
|---|---|
| Code (data gen + training + eval) | github.com/hamishivi/tmax |
| Models (TMax-4B, TMax-9B) | HF: allenai/tmax |
| Dataset (TMax-15K-Harbor) | Harbor registry |
| Paper | arXiv 2606.23321 |
License: Apache 2.0.
Why This Matters
RL training for terminal agents had no reproducible open baseline. TMax closes that gap — and it does so without scaling tricks: 9B parameters, simple outcome-only RL, a data generation design that makes corpus composition a tunable parameter rather than a fixed scrape artifact.
The compositional sampler approach is worth borrowing beyond this specific project: treating tasks as products of orthogonal axes gives you explicit control over difficulty distribution and task-type coverage, which is exactly what you need when training with RL.
For teams working on terminal agents, TMax is the strongest available starting point: reproducible benchmark data, scalable data generation, open training recipe, and open model weights for comparison.
Repository: github.com/hamishivi/tmax · Paper: arXiv 2606.23321 · Models: allenai/tmax
Mycelium Protocol — tracking the deep evolution of AI systems
© 2026 Mycelium Protocol. All rights reserved.
关于本站 · 免责声明
🍄 Mushroom Research Blog 是非营利、免费公开的个人科技观察博客与公众号 XStack18,不接受商业合作、不代表任何企业或机构立场,也不谋求商业利益。我们以个人视角客观中立地记录和分析 AI、Web3 等领域的最新模型发布与技术动态——不止转述新闻标题或二手信息,而是给出有独立思考的深入分析,希望帮更多人获得有价值的一手科技认知。
⚠️ 文中介绍的开源代码与模型,仅供学习交流与技术借鉴。它们大多仍处于早期阶段,有待进一步研究和验证,请勿直接用于工作或生产环境;如需采用,请先自行充分测试,并核实其许可证与安全性。
Open-source code and models featured here are shared for learning and reference only. Most are early-stage and still need further study and verification — please don't use them directly in your work or in production. Test them thoroughly and check their licenses and security first.
- 本站文章均为作者基于公开信息的个人研究与观点整理,不代表文中提及的任何公司、产品、模型的官方立场,未与其构成商业关联或合作关系。
- 科技行业信息更新极快,我们尽力保证内容准确、及时,但不对完整性、实时性做绝对保证,具体请以相关企业/项目官方公告为准。
- 文中引用的第三方商标、产品名称、图片、数据等版权归原权利人所有,我们会尽量注明来源;如你认为存在版权疑问或侵权,请通过下方邮箱联系我们,收到通知后会尽快核实处理(更正、加注来源或删除)。
- 文章内容仅为技术科普与个人观点,不构成投资、法律或其他专业建议,据此进行任何决策的后果需自行判断和承担。
📮 侵权 / 勘误 / 合作咨询:[email protected]
💬 评论与讨论
使用 GitHub 账号登录后发表评论