跳转至

T1:面向长程终端任务的强化学习智能体

T1: Terminal Agent Reinforcement Learning for Long-Horizon Tasks

arXiv:2609.11042 [cs.LG]
Subjects: Machine Learning (cs.LG), Artificial Intelligence (cs.AI)
Authors: Junyao Yang, Yucheng Shi, Zhongzhi Li, Ruhan Wang, Zongxia Li, Haitao Mi, Leowei Liang
Submitted: 10 September 2026
arXiv ID: arXiv:2609.11042
DOI: 10.48550/arXiv.2609.11042

文章背景与核心概要

随着 AI 智能体 (AI Agent) 的应用场景加速拓展至自动化编程、科学探索等长链条复杂场景,在命令行终端中稳定执行长程任务成为了核心技术制高点。传统大模型在面对需要数百步工具调用的任务时,往往容易发生状态漂移、轨迹发散和过拟合。本项研究推出了总参数量达 122B 的混合专家模型 (Mixture-of-Experts, MoE) 智能体——T1。该模型通过在云端沙箱中操控真实 Shell 终端并依托强化学习进行端到端训练,结合密集过程奖励、Token 标识符训练优化 (TITO) 与路由重放 (R3) 等创新训练配方,在 Terminal-Bench 2.1 上取得了 64.0% 的高解题率,并在超长程终端基准测试中以 27.9% 的成绩超越了 GPT-5.4 与 GLM-5.1。


📋 内容摘要

Summary

T1 是一个总参数量达 122B 的混合专家模型 (Mixture-of-Experts, MoE) ,通过强化学习 (Reinforcement Learning) 训练以在云端沙盒环境中执行长程终端任务(单任务可支持处理超过 300 轮以上的工具调用)。为了实现稳定的行动者-评论家 (Actor-Critic) 训练,并获得真实的通用能力迁移而非在基准测试上过拟合,作者团队提出了一套全面的训练配方:包含密集过程奖励机制、Token 标识符训练优化 (Token Identifiers Training Optimization, TITO) 以及 Rollout 路由重放 (Rollout Routing Replay, R3)。在 Terminal-Bench 2.1 基准测试中,T1 成功解决了 64.0% 的任务(相较基座模型的 43.8% 实现了大幅跃升);而在更具挑战性的长程终端基准 Long-Horizon Terminal Bench 上,其解题率达到 27.9% ,表现超越了 GPT-5.4 与 GLM-5.1 等前沿模型。

T1 is a 122B-parameter Mixture-of-Experts (MoE) model trained via reinforcement learning to execute long-horizon terminal tasks in a cloud sandbox environment (handling over 300+ tool-call turns per task). To achieve stable actor-critic training and genuine capability transfer without benchmark overfitting, the authors introduce a comprehensive training recipe featuring dense process rewards, Token Identifiers Training Optimization (TITO), and Rollout Routing Replay (R3). On Terminal-Bench 2.1, T1 successfully resolves 64.0% of tasks (up from 43.8% in the base model), and on the Long-Horizon Terminal Bench, it reaches 27.9%, outperforming models like GPT-5.4 and GLM-5.1.


📑 论文摘要

Abstract

智能体的实际应用正在快速转向涉及长决策链的复杂任务(如自主编程与科学探索),其中命令行终端操作类任务显得尤为关键。我们隆重推出了 T1——这是一个总参数规模达 122B、采用强化学习进行端到端训练的混合专家模型。它能在云端隔离沙箱中操控真实的 Shell 环境,在单个长程任务中连续执行超过 300 轮以上的工具调用,并通过运行每项任务专属的验证器程序来获取奖励反馈信号。

Agent usage is shifting toward long-horizon tasks such as coding and scientific discovery, among which terminal tasks are especially important. We introduce T1, a Mixture-of-Experts model of 122B total trained with reinforcement learning, operating a real shell in a cloud sandbox for up to 300+ tool-call turns per task, rewarded by executing each task's own verifier.

我们为此构建了一套完整的工程与训练配方: 1. 强力热启动预热:通过密集过程奖励机制对完整交互轨迹打分(依据通过验证器的绝对数量),从而有效稳定 Actor-Critic 强化学习训练过程。 2. 基于 TITO 与 R3 的高稳定性优化: - TITO 构建:基于采样时实际生成的精确 Token 标识符进行训练,并在多轮交互的轮次边界处引入漂移修复机制。 - Rollout 路由重放 (R3):在 MoE 的每一个专家层中,严格记录采样器在每个 Token 上的专家选择路径,并在梯度反向传播训练期间无缝重放。 3. 分布外 (OOD) 强化训练语料:构建与 Terminal-Bench 2.1 彻底隔离的独立种子与合成任务池,以确保模型性能的飞跃体现的是真实的跨领域能力迁移,而非针对评估基准的单纯过拟合。

We provide a comprehensive recipe: 1. Aggressive Warm-Starting: Stabilizes actor-critic training, utilizing a dense process reward scoring trajectories by the absolute number of passing verifiers. 2. Stable Optimization via TITO and R3: - TITO Construction: Training on the exact sampled token identifiers with drift repair at turn boundaries. - Rollout Routing Replay (R3): Recording the sampler's per-token expert choices at every MoE layer and replaying them during training. 3. Out-of-Distribution Training Corpus: Isolated seeds and synthesized tasks disjoint from Terminal-Bench 2.1 to ensure gains reflect genuine capability transfer over benchmark overfitting.

通过 TITO 与 R3 的协同优化,我们将模型训练与推理阶段的对数概率差异从 0.021 骤降至 0.013 ,并在损失计算区域实现了完全对齐、零 Token 漂移。

Together, TITO and R3 cut the training-to-inference log-probability difference from 0.021 to 0.013, with exactly aligned zero token drift in the loss region.


📊 评测结果与模型表现

Results & Performance

  • Terminal-Bench 2.1 基准:后训练流水线将模型初始解题率从基座模型的 43.8% 提升至 64.0%
  • Long-Horizon Terminal Bench 长程基准:T1 达到了 27.9% 的通过率,成功超越了 GPT-5.4 与 GLM-5.1 。
  • Terminal-Bench 2.1: Post-training pipeline raises the initial base model performance from 43.8% to 64.0% resolved.
  • Long-Horizon Terminal Bench: T1 reaches 27.9%, successfully surpassing GPT-5.4 and GLM-5.1.

🔗 全文阅读与相关资源

Full-Text & Resources

license icon