[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-49d0aaf8-6fcf-4bf9-83fb-19a016ae2784":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":23,"created_at":24,"modified_at":25,"is_published":26,"publish_type":27,"image_url":13,"view_count":28},"49d0aaf8-6fcf-4bf9-83fb-19a016ae2784","CompactionRL:把上下文压缩塞进 RL 循环,GLM-5.2 训练管线吃下 5–7pp 编码代理增益","智谱 (Z.ai) 团队 7 月 6 日在 arXiv 公开 CompactionRL (2607.05378),把「上下文压缩」从推理期招数重写为可训练的 RL 原语,直接服务 GLM-5.2 (750B-A40B) 的训练管线。方法用跨轨迹 GAE + token 级 loss 归一化,把任务执行与摘要生成联合优化,让模型在同一份策略里既会干活又会压缩。\n\n效果在开源 MoE 上复现得很整齐:GLM-4.5-Air (106B-A30B) SWE-bench Verified Pass@1 拉到 66.8% (+7.0)、Terminal-Bench 2.0 24.5% (+3.1);GLM-4.7-Flash (30B-A3B) 同两项分别 +5.5、+6.8,跑出 56.0% \u002F 20.2%。\n\n更值得注意的是其解决的真问题:长程 agent 轨迹一旦被压缩,GRPO 这类基于 group-level advantage 的方法会失效,因为同 prompt 不同 rollout 的子轨迹数与长度不再对齐;CompactionRL 切到 critic-based PPO + token 级优势估计,把 RL 的训练假设从「固定长度轨迹」改成「可变长子轨迹」。\n\n工程上也给出基座:训练\u002Frollout 用智谱自研 slime 框架,支持 white\u002Fblack-box rollout、sub-agent 工作流与 FP8 KV-cache,正好承载 GLM-5.2 的并行 OPD 后训练负载。这意味着 RL 流水线首次把「上下文预算」当成可学习的旋钮,而不只是推理时的兜底——对所有还在纠结 agent 长程规划的团队,这是一份可复用的训练配方。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.05378","1eab5c4a-0c8e-49a4-8ac8-0f84a2a3c3a4",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"e82b2d09-81b2-43d1-977e-e018443b3c14","coding-agent",{"id":18,"name":19,"slug":19,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source","2026-07-10T12:08:00Z","2026-07-10T12:08:20.972757Z","2026-07-10T12:08:20.972766Z",true,"agent",3]