[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-glm-5-3-post-training-stack-deep-dive":3,"news-related-6b203495-fcab-4afe-baa7-1079cf993796":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},"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 的基座与 GLM-5.2 完全相同(743B MoE),官方博客直言本次全部提升来自后训练扩展。本文拆解其三件套:IndexShare 长上下文处理、SAO 长程任务 RL、slime 异步训练框架,以及研究代理造环境、裁判代理验可解、合成验证器产出二元奖励的完整流水线,和训练-推理 logprob 对齐控制在 1e-7、长程编码 RL 吞吐提升 2.3 倍的系统工程细节。","# 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\n\n8 月 14 日,Z.ai 发布 GLM-5.3。这条新闻最反常规的地方不是跑分,而是官方博客的开场白——**「Scaling post-training is all we did for GLM-5.3」**(对 GLM-5.3,我们做的只有扩展后训练)。基座与 GLM-5.2 完全相同:同样的 743B MoE,一个参数都没改,所有能力提升都来自后训练阶段。\n\n## 冻结基座,是一笔工程账\n\n重新预训练一个数百亿级模型需要数万 GPU、数月时间和巨额算力投入;而冻结基座意味着下游推理基础设施、算子优化可以零成本迁移,社区在上一代模型上的微调工作也完全兼容。当预训练参数规模的边际回报收窄,竞争重心就从「谁的参数多」转向「谁在后训练阶段训练得更好」。\n\n支撑这条路线的是一个三件套技术栈(官方博客附了三个链接):\n\n- **IndexShare**(论文 [arXiv:2603.12201](https:\u002F\u002Farxiv.org\u002Fabs\u002F2603.12201)):高效长上下文处理;\n- **SAO**(论文 [arXiv:2607.07508](https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.07508)):面向长程任务的 RL 算法;\n- **slime**(开源仓库 [THUDM\u002Fslime](https:\u002F\u002Fgithub.com\u002FTHUDM\u002Fslime)):大规模异步训练框架,训练侧用 Megatron、rollout 侧用 SGLang。\n\n## 真正的瓶颈:不是模型,是环境\n\n官方博客里最有信息量的一句话是:**随着 agent 能力提升,扩展后训练的难度大量从模型转移到环境**。一个有用的任务环境必须可执行、可验证、贴近真实专业工作——而且需要很多个,不是几个手工搭建的样本。\n\nZ.ai 的解法是一条端到端的环境合成流水线:\n\n1. **Research Agent** 从真实工作中收集任务模式,转化为带多步依赖和隐藏状态的可运行长程环境;\n2. **Judge Agent** 逐一试做这些任务,验证它们真的可解;\n3. **合成验证器**——注意,验证器在合成时接触不到参考答案;\n4. 用求解轨迹去发现并关闭奖励捷径(reward shortcut);\n5. 通过 oracle、no-op、unsolved-state 三重检查后,产出可直接用于训练的二元奖励。\n\n这套流水线目前仍需要可观的人在环参与,官方也承认「让环境生成和验证更自主」是下一步。\n\n## slime:训练系统里的硬功夫\n\n所有训练跑在 slime 上。它的设计把训练、rollout 和数据缓冲统一在一条数据流里——数学、代码、沙箱、验证器和长程 agentic 环境**以数据生成的方式插入,而不是对训练循环的修改**。这就是 GLM-5.2 到 5.3 能持续加环境、不用重建训练栈的原因。\n\n本次迭代在算法侧加了 top-p mask、top-k 和全词表 OPD,以及提升训练-rollout 一致性的配置。最硬的数字是:**训练与 rollout 两侧的平均 logprob 差异控制在 1e-7 量级,比之前的配置降低超过 99.99%**。系统侧,本地存储成为额外缓存层,多教师 OPD 支持动态切换加预取——不用为每个教师常驻一个推理服务;再加上 router 与 slime 的联合调度和负载均衡,官方称**长程编码 RL 的端到端训练吞吐提升超过 2.3 倍**。\n\n## 效果:跑分之外,看 token 账单\n\n结果层面:Terminal-Bench 3.0 从 4.6 涨到 28.3,DeepSWE v1.1 从 46.2 到 66.9,CyberGym 从 77.2% 到 84.5%(官方称超过 Mythos 5 的 83.8% 与 GPT-5.6 Sol 的 83.6%)。更值得看的是自建 Z.ai Code Bench 上的 token 效率:High 档 GLM-5.3 用约 50K 输出 token 达到 31.4% 准确率,超过 Claude Opus 4.8 用 120K token 达到的 29.5%——一半不到的成本,更好的结果。\n\n真实世界同样有验证:与中国多个安全团队合作,经专家评审、筛选、去重后,模型在 269 个项目中识别出 2,436 个漏洞(107 个 Critical、990 个 High),最老的漏洞可追溯到 1981 年,平均已在代码库中存活 26.6 年,全部记录在公开台账 [cvd.z.ai](https:\u002F\u002Fcvd.z.ai\u002F)。权重将在发布两周后、完成安全评估与加固后公开。\n\n## 所以呢\n\nGLM-5.3 给行业的信号很直接:当基座参数规模收敛,「谁的环境工厂造得好、谁的 RL 系统跑得稳」成为新的分水岭。环境合成流水线 + 1e-7 级训练一致性 + 2.3 倍吞吐——这些不如跑分性感,却决定了下一次能力跃迁的成本和速度。下一次看到某家模型「突然变强」,值得先问一句:它的环境是怎么造出来的?\n\n> 素材来源:[Z.ai 官方博客](https:\u002F\u002Fz.ai\u002Fblog\u002Fglm-5.3)、[HappyRock 深拆](https:\u002F\u002Fhappyrock.cloud\u002Fblog\u002F2026-08-16_glm-5.3_en\u002F)(2026-08-16)","https:\u002F\u002Fhappyrock.cloud\u002Fblog\u002F2026-08-16_glm-5.3_en\u002F","df9ef325-77c5-4e95-9c03-f6cf5b150ef0",[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},"7ac06d8e-b074-4147-abfc-ffaa4c6b8744","ai-efficiency",{"id":19,"name":20,"slug":20,"description":14,"color":14},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",{"id":22,"name":23,"slug":23,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"f01019f5-9aa4-479a-b1af-9e17732cbe5f","en","Inside GLM-5.3's post-training factory: base untouched","GLM-5.3 shares its base model with GLM-5.2 (743B MoE, unchanged), and Z.ai's official blog states that every gain comes from post-training scaling. This piece breaks down the three-part stack — IndexShare for long-context, SAO for long-horizon RL, and the slime asynchronous training framework — plus the full pipeline where research agents synthesize environments, judge agents verify solvability, and synthesized verifiers emit binary rewards, along with the engineering details: training-rollout logprob alignment at 1e-7 and a 2.3x throughput gain on long-horizon coding RL.","# Inside GLM-5.3's Post-Training Factory: A Frozen Base, Environment Synthesis, and 1e-7 Alignment Carrying Every Gain\n\nOn August 14, Z.ai released GLM-5.3. The most unconventional thing about this launch is not the benchmark numbers but the opening line of the official blog — **\"Scaling post-training is all we did for GLM-5.3.\"** The base model is identical to GLM-5.2: the same 743B MoE, not a single parameter changed, with every capability gain coming from the post-training phase.\n\n## Freezing the Base Is an Engineering Decision\n\nRe-pretraining a hundred-billion-class model takes tens of thousands of GPUs, months of time, and massive compute investment; freezing the base means downstream inference infrastructure and operator optimizations migrate at zero cost, and the community's fine-tuning work on the previous generation remains fully compatible. As marginal returns on pretraining parameter scale narrow, the battleground shifts from \"who has more parameters\" to \"who trains better in post-training.\"\n\nSupporting this route is a three-part stack (the official blog links all three):\n\n- **IndexShare** (paper [arXiv:2603.12201](https:\u002F\u002Farxiv.org\u002Fabs\u002F2603.12201)): efficient long-context processing;\n- **SAO** (paper [arXiv:2607.07508](https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.07508)): RL for long-horizon tasks;\n- **slime** (open-source repo [THUDM\u002Fslime](https:\u002F\u002Fgithub.com\u002FTHUDM\u002Fslime)): a large-scale asynchronous training framework, with Megatron on the training side and SGLang on the rollout side.\n\n## The Real Bottleneck: Not the Model, the Environments\n\nThe most information-dense sentence in the official blog is this: **as agent capability improves, much of the difficulty in scaling post-training moves from the model to the environment.** A useful task environment has to be executable, verifiable, and close to real professional work — and you need many of them, not a handful of hand-built samples.\n\nZ.ai's answer is an end-to-end environment synthesis pipeline:\n\n1. A **Research Agent** collects task patterns from real work and turns them into runnable long-horizon environments with multi-step dependencies and hidden state;\n2. A **Judge Agent** attempts each task to verify it is actually solvable;\n3. **Verifiers are synthesized** — note that the verifier has no access to the reference solution during synthesis;\n4. Solver trajectories are used to discover and close reward shortcuts;\n5. After passing oracle, no-op, and unsolved-state checks, the verifier produces a binary reward reliable enough to train on directly.\n\nThe pipeline still requires a meaningful amount of human-in-the-loop work; the team admits making environment generation and verification more autonomous is one of the next steps.\n\n## slime: The Hard Engineering Inside the Training System\n\nAll training runs on slime. Its design keeps training, rollout, and the data buffer on a single dataflow — math, code, sandboxes, verifiers, and long-horizon agentic environments plug in **as data generation rather than as changes to the training loop**. That is what let Z.ai keep adding environments from GLM-5.2 through 5.3 without rebuilding the training stack.\n\nThis iteration added top-p mask, top-k and full-vocabulary OPD on the algorithmic side, plus configurations improving training-rollout consistency. The hardest number: **the average logprob difference between the training and rollout sides is controlled at the 1e-7 level, a reduction of more than 99.99% versus previous setups.** On the systems side, local storage serves as an additional caching layer; multi-teacher OPD supports dynamic teacher switching with prefetching — no dedicated long-running inference service per teacher — and with joint scheduling and load balancing between the router and slime, the team reports **end-to-end RL training throughput on long-horizon coding improved by more than 2.3x**.\n\n## Results: Beyond Benchmarks, Read the Token Bill\n\nOn results: Terminal-Bench 3.0 rose from 4.6 to 28.3, DeepSWE v1.1 from 46.2 to 66.9, and CyberGym from 77.2% to 84.5% (ahead of Mythos 5's 83.8% and GPT-5.6 Sol's 83.6%, per the official post). More telling is token efficiency on the in-house Z.ai Code Bench: at High effort, GLM-5.3 reaches 31.4% accuracy at roughly 50K output tokens, surpassing Claude Opus 4.8's 29.5% at 120K tokens — better results at less than half the cost.\n\nThere is real-world validation too: working with several security teams in China, after expert review, screening, and deduplication, the model identified 2,436 vulnerabilities across 269 projects (107 Critical, 990 High), with the oldest flaw dating back to 1981 and the average vulnerability having lived 26.6 years in the codebase — all tracked in the public ledger at [cvd.z.ai](https:\u002F\u002Fcvd.z.ai\u002F). Weights will be released two weeks after launch, once safety evaluation and hardening are complete.\n\n## So What\n\nThe signal GLM-5.3 sends to the industry is direct: as base-model parameter scales converge, \"whose environment factory is built better and whose RL system runs more stably\" becomes the new dividing line. An environment synthesis pipeline + 1e-7-level training consistency + 2.3x throughput — less sexy than benchmark scores, yet they determine the cost and speed of the next capability jump. The next time a model \"suddenly gets stronger,\" it is worth asking first: how were its environments built?\n\n> Sources: [Z.ai official blog](https:\u002F\u002Fz.ai\u002Fblog\u002Fglm-5.3), [HappyRock deep dive](https:\u002F\u002Fhappyrock.cloud\u002Fblog\u002F2026-08-16_glm-5.3_en\u002F) (2026-08-16)","glm-5-3-post-training-stack-deep-dive","2026-08-17T13:00:00Z","2026-08-16T19:14:58.036503Z","2026-08-16T19:14:58.036514Z",true,"agent",288,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"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":46,"title":47,"news_slug":48,"published_at":49},"259d91b2-ed6b-4af8-8f2a-f759b84cc617","蚂蚁 Ling-3.0 Flash：124B\u002F5.1B MoE 的 Agent 生产级模型","inclusionai-ling-3-flash-hybrid-linear-moe-agent","2026-08-14T08:00:00+00:00",{"id":51,"title":52,"news_slug":53,"published_at":54},"ed39ed38-b5fa-4f58-92cf-d05233ab998b","Speculate with Memory：LLM Agent 无损加速 2.5×，准确率涨 39pp","speculate-with-memory-2-5x","2026-07-15T08:15:00+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"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",{"id":61,"title":62,"news_slug":63,"published_at":64},"c5fb2ca8-891d-4c8c-96bb-847a78a48255","OmniAgent 把视频理解变成「主动感知」：Qwen 团队 7B 全模态代理跑赢 72B「看完全片」","omniagent-qwen-active-perception-7b-vs-72b","2026-06-21T12:01:00+00:00",{"id":66,"title":67,"news_slug":68,"published_at":69},"e3d0e837-43fc-49a6-b533-fbb75aff95d3","FastContext 4B-30B 子代理：把 60% 推理 Token 留给\"找代码\"，Coding Agent 提速 5.5%","fastcontext-msr-sub-agent-code-search-5-5pct","2026-06-20T02:01:00+00:00"]