[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-longstraw-2-1m-token-rl":3,"news-related-582462cb-d952-4136-818b-3edd6fc9ad0c":33},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":20,"news_slug":26,"published_at":27,"created_at":28,"modified_at":29,"is_published":30,"publish_type":31,"image_url":13,"view_count":32},"582462cb-d952-4136-818b-3edd6fc9ad0c","LongStraw 把百万 token 拉进 RL 后训练:8 张 H20 跑通 2.1M token 端到端","**核心问题。** 现在前沿推理系统已经能吞百万 token context，但 RL 后训练仍卡在 256K——把训练上下文从推理扩到 2M+，是 agentic RL 真正的瓶颈。论文观察到，AI agents 的 observation、tool output、document、prior decision 沿 trajectory 累积，post-training 长度上不去，长程 agent 能力就出不来。\n\n**LongStraw 的做法。** 2026-07-16 上 arXiv（编号 2607.14952，MindLab-Research，Changhai Zhou 等 20 位作者）给出了一个架构感知的执行栈，专为百万 token RL post-training 设计。三个关键动作：(1) 共享 prompt 不开 autograd，inference-only 评估前缀，只保留后续 token 真正需要的模型态；(2) response 短分支逐次回放，把长 response 拆成短分支一段一段 replay，把\"活的训练图\"砍到最小；(3) 架构专属优化，不写通用解法，而是针对 hybrid recurrent + full-attention 的 Qwen3.6-27B 和 compressed-attention MoE 的 GLM-5.2 两个真实基座做特化。算法侧用的是当下最流行的 GRPO。\n\n**实测数据。** 8 张 H20 上，Qwen3.6-27B 的 grouped scoring + response backward 可以跑到 2.1M positions，group size 从 2 扩到 8 只多占用 **0.21 GB** 峰值显存；极限压测跑到 4.46M。32 张 H20 上，GLM-5.2 全部 78 层端到端 forward 在 2.1M token prompt 下被跑通——这条含金量更高，因为 compressed-attention MoE 的层间状态远更复杂。\n\n**个人观点。** 这条工作关键不在\"训完一个完整 RL 步\"，而在\"敢把这套执行栈如实放出\"。作者直接标注 \"establishes execution capacity rather than complete training correctness, some distributed forward and gradient composition paths remain incomplete\"——这种克制在 RL infra 论文里太少见了，比\"又一个 1M context RL\"的套路话踏实得多。agentic RL 的真实瓶颈就在上下文长度，LongStraw 把\"百万 token 后训练\"从口号变成可执行的工程件，代码已开源在 github.com\u002FMindLab-Research\u002Flongstraw，下一步就看谁能用它做出真正能训出 long-horizon agent 的完整训练流程。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.14952","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17],{"id":11,"name":12,"slug":12,"description":13,"color":13},"7ac06d8e-b074-4147-abfc-ffaa4c6b8744","ai-efficiency",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":18,"name":19,"slug":19,"description":13,"color":13},"4f214978-cac1-4f39-aa4b-f92a0d0934b7","transformer",[21],{"id":22,"lang":23,"title":24,"summary":25,"content":25},"e9e1292a-e2eb-4449-bbc1-aeb4d730f943","en","LongStraw brings million-token sequences into RL training","**Core problem.** Frontier reasoning systems can now swallow million-token contexts, but RL post-training is still stuck at 256K — pushing training context from inference to 2M+ is the real bottleneck for agentic RL. The paper observes that AI agents' observations, tool output, documents, and prior decisions accumulate along the trajectory; if post-training length can't go up, long-horizon agent capabilities can't come out. **What LongStraw does.** arXiv 2607.14952 (MindLab-Research, Changhai Zhou and 20 co-authors), posted 2026-07-16, gives an architecture-aware execution stack designed for million-token RL post-training. Three key moves: (1) shared prompt doesn't open autograd, inference-only evaluation prefix, keeping only the model state truly needed by subsequent tokens; (2) response short-branch replay, splitting long responses into short branches replayed segment by segment, compressing the \"live training graph\" to the minimum; (3) architecture-specific optimization — instead of writing a general solution, specialize for two real backbones: hybrid recurrent + full-attention Qwen3.6-27B, and compressed-attention MoE GLM-5.2. The algorithm side uses today's most popular GRPO. **Measured data.** On 8 H20s, Qwen3.6-27B's grouped scoring + response backward runs to 2.1M positions, with group size going from 2 to 8 only adding **0.21 GB** of peak VRAM; extreme stress test reached 4.46M. On 32 H20s, GLM-5.2's all 78 layers of end-to-end forward were run through on 2.1M-token prompts — this carries more weight because compressed-attention MoE's inter-layer state is far more complex. **Personal take.** The key here isn't \"finishing a complete RL step\", but \"daring to release this execution stack as-is\". The authors explicitly mark \"establishes execution capacity rather than complete training correctness, some distributed forward and gradient composition paths remain incomplete\" — this restraint is too rare in RL infra papers, far more solid than the cliché talk of \"another 1M context RL\". The real bottleneck of agentic RL is right at context length; LongStraw turns \"million-token post-training\" from a slogan into executable engineering. Code is open-sourced at github.com\u002FMindLab-Research\u002Flongstraw. The next step is who can use it to make a complete training flow that truly trains long-horizon agents.","longstraw-2-1m-token-rl","2026-07-16T13:00:00Z","2026-07-17T20:09:48.312168Z","2026-08-19T02:08:40.142862Z",true,"agent",116,{"items":34},[35,40,45,50,55,60],{"id":36,"title":37,"news_slug":38,"published_at":39},"e3f049f5-2f0d-48d2-8e88-246ef006fa16","LoopMTP 给循环 Transformer 装上前瞻路标：固定参数下让每一轮都做不同的事","loopmtp-latent-multi-token-loop-guidance","2026-08-04T13:13:09+00:00",{"id":41,"title":42,"news_slug":43,"published_at":44},"c48681ff-ddbb-402c-ade8-23b584a06aea","更强教师反而教不动学生：Lightning OPD 2.0 剥掉蒸馏中的“文风噪声”","lightning-opd-2-cross-teacher-style-bias","2026-07-30T16:17:15+00:00",{"id":46,"title":47,"news_slug":48,"published_at":49},"a151db0c-d832-4df2-ac03-2d4e58b26e99","Kimi K3 跑通 MiniTriton:Moonshot 让 LLM 第一次从零编译出自己的 GPU 编译器","kimi-k3-minitriton-gpu-compiler","2026-07-26T14:00:00+00:00",{"id":51,"title":52,"news_slug":53,"published_at":54},"d132e369-59e6-40a6-9d20-0c5d94535026","Kaplan scaling law 原论文有 bug:原作者 ICML 2026 公开承认,行业被'大而不精'拐了两年","kaplan-scaling-law-bug","2026-07-05T15:00:00+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"7312935f-7432-4dbd-9cff-c665fa7b4765","FlashMorph:ByteDance Seed 把混合注意力的\"层选择\"做成预算约束优化","flashmorph-hybrid-attention-layer","2026-07-05T02:01:00+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"b47914fe-d539-440f-ab02-3b089120051f","MultiHashFormer 用多重哈希签名改写 LLM 词表嵌入：3B 模型跑赢标准 Transformer","multihashformer-hash-embedding-3b","2026-06-29T04:20:00+00:00"]