GLM-5：从氛围编码到智能体工程

GLM-5: from Vibe Coding to Agentic Engineering

唐杰 Jie Tang · Zhipu AI · 2026-02-17 · arXiv:2602.15763 ↗ · 被引 241

摘要 · Abstract

我们提出了 GLM-5，这是一个下一代基础模型，旨在将氛围编码范式转变为智能体工程。在其前身智能体、推理和编码能力的基础上，GLM-5 采用 DSA 显著降低训练和推理成本，同时保持长上下文保真度。为了推进模型对齐和自主性，我们实现了一种新的异步强化学习基础设施，通过将生成与训练解耦，大幅提升后训练效率。此外，我们提出了新颖的异步智能体强化学习算法，进一步提高了强化学习质量，使模型能够更有效地从复杂的长期交互中学习。通过这些创新，GLM-5 在主要开放基准测试中达到了最先进的性能。最关键的是，GLM-5 在真实世界编码任务中展现了前所未有的能力，在处理端到端软件工程挑战方面超越了之前的基线。代码、模型和更多信息可在 https://github.com/zai-org/GLM-5 获取。

We present GLM-5, a next-generation foundation model designed to transition the paradigm of vibe coding to agentic engineering. Building upon the agentic, reasoning, and coding (ARC) capabilities of its predecessor, GLM-5 adopts DSA to significantly reduce training and inference costs while maintaining long-context fidelity. To advance model alignment and autonomy, we implement a new asynchronous reinforcement learning infrastructure that drastically improves post-training efficiency by decoupling generation from training. Furthermore, we propose novel asynchronous agent RL algorithms that further improve RL quality, enabling the model to learn from complex, long-horizon interactions more effectively. Through these innovations, GLM-5 achieves state-of-the-art performance on major open benchmarks. Most critically, GLM-5 demonstrates unprecedented capability in real-world coding tasks, surpassing previous baselines in handling end-to-end software engineering challenges. Code, models, and more information are available at https://github.com/zai-org/GLM-5.

核心贡献 · Key contributions

提出 GLM-5，一个 744B 参数的基座模型，实现从 vibe coding 到智能体工程的范式转变。
Proposes GLM-5, a 744B-parameter foundation model transitioning from vibe coding to agentic engineering.
采用 DSA（DeepSeek 稀疏注意力）降低训练和推理成本，同时保持长上下文保真度。
Adopts DSA (DeepSeek Sparse Attention) to reduce training and inference costs while maintaining long-context fidelity.
实现异步强化学习基础设施，将生成与训练解耦，提升后训练效率。
Implements asynchronous reinforcement learning infrastructure decoupling generation from training for efficient post-training.
引入新型异步智能体强化学习算法，从复杂长程交互中学习。
Introduces novel asynchronous agent RL algorithms for learning from complex, long-horizon interactions.
在主要开放基准和真实编码任务上达到最先进性能。
Achieves state-of-the-art performance on major open benchmarks and real-world coding tasks.
在七个国产芯片平台上实现全栈适配中国 GPU 生态系统。
Full-stack adaptation to Chinese GPU ecosystems across seven domestic chip platforms.

局限 · Limitations

DSA 训练需要两阶段预热和稀疏适应，增加了复杂性。
DSA training requires a two-stage warm-up and sparse adaptation, adding complexity.
异步强化学习引入离策略偏差，需要仔细的裁剪和过滤。
Asynchronous RL introduces off-policy bias requiring careful clipping and filtering.
智能体环境的可扩展性限于 1 万多个真实软件工程任务。
Scalability of agentic environments is limited to 10K+ real-world SWE tasks.
强化学习期间冻结 DSA 索引器可能限制动态环境中的适应能力。
DSA indexer freezing during RL may limit adaptation in dynamic settings.
评估聚焦于开放基准，真实部署挑战仍然存在。
Evaluation focuses on open benchmarks; real-world deployment challenges remain.

论文章节 · Sections（共 26）

摘要 Abstract
1 引言 1 Introduction
2 预训练 2 Pre-Training
2.1 架构 2.1 Architecture
2.2 预训练数据 2.2 Pre-training Data
2.3 中期训练 2.3 Mid-Training
2.4 训练基础设施 2.4 Training Infrastructure
3 后训练 3 Post-Training
3.1 监督微调 3.1 Supervised Fine-Tuning
3.2 推理强化学习 3.2 Reasoning RL
3.3 智能体强化学习 3.3 Agentic RL
3.4 通用强化学习 3.4 General RL
3.5 同策略跨阶段蒸馏 3.5 On-Policy Cross-Stage Distillation
3.6 强化学习训练基础设施：slime 框架 3.6 RL Training Infrastructure: The slime Framework
4 智能体工程 4 Agentic Engineering
4.1 智能体任务的异步强化学习 4.1 Asynchronous RL for Agentic Tasks
4.2 智能体的环境扩展 4.2 Environment Scaling for Agents
5 将 GLM-5 适配至中国芯片基础设施 5 Adapting GLM-5 to Chinese Chip Infrastructure
6 评估 6 Evaluation
6.1 ARC 基准评估 6.1 Evaluation of ARC Benchmarks
6.2 真实世界智能体工程经验评估 6.2 Evaluation of Real-world Agentic Engineering Experience
6.3 真实世界通用能力评估 6.3 Evaluation of Real-world General Abilities
7 结论 7 Conclusion
8 彩蛋 8 Easter Eggs
9 贡献 9 Contribution
致谢 Acknowledgement

阅读逐段中英对照全文 →