[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-alibaba-qwen-qwengyre-elastic-rl-xlong-agents":3,"topics-all":38,"news-related-9fcb5d74-315e-4d57-a7da-290fbfb9e91c":57},{"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},"9fcb5d74-315e-4d57-a7da-290fbfb9e91c","阿里 Qwen 发布弹性 RL 框架 QwenGyre,训练长时 Agent 提速 1.85x","阿里 Qwen 团队发布 QwenGyre 弹性强化学习框架,专为小时级超长 Agent 训练设计。该框架通过弹性调度 GPU 在 rollout 与训练间切换,并将分支轨迹转化为有界训练样本,在 Qwen 3.8 2.4T 模型上把 NL2RepoBench 任务通过率从 52.5% 提升到 58.5%。","把一个编码 Agent 扔进真实的软件工程任务里跑,经常意味着一次 rollout 就要消耗多小时、上百次模型与工具交互、滚出接近 1M tokens。这种「超长时长」(xlong-horizon)的训练,正在成为 Agent RL 的核心瓶颈。GPU 要么在等最慢的那次 rollout,要么在堆积的空泡里闲置;而当 Agent 像 Claude Code 那样压缩上下文、生成子 Agent、改方向重试时,一次执行会变成一棵分支共享前缀的轨迹树,展开来训又会把显存与计算量撑爆。阿里 Qwen 团队最新放出的 QwenGyre,就是冲着这两个问题来的。\n\n## 弹性调度:硬件跟着活儿走,而不是反过来\n\nQwenGyre 把多机 GPU 切成若干「单元」(cell),每个单元既能在 SGLang 上服务 rollout,也能在 Megatron 里跑一次完整的训练步。核心单元持有权威权重和优化器状态,卫星单元平时只服务 rollout,等调度器判定「剩下的单元仍然够撑住现有请求」,就立刻往训练角色切换。\n\n切换不需要 Agent 配合——所有模型调用都打在一个 proxy 后面,proxy 通过 RDMA 把 KV-cache block 顺手搬走,KV 复用最大化、重算前缀被压到最低。Mooncake 的 Transfer Engine 让卫星单元把核心权重拉过来直接加入正在进行的训练 batch,流式训练再让单元微步级别地领取数据。实测一次角色切换平均只需 8.52 秒(rollout→训练)和 3.46 秒(训练→rollout)。\n\n这套设计直接体现在端到端速度上。在匹配 GPU 预算与延迟的前提下,NL2RepoBench、DeepSWE、TerminalBench 三个基准都跑过 Qwen 3.6 122B 与 Qwen 3.8 2.4T 两套模型:对比 Async 模式提速 1.38x–1.78x,对比 Colocate 模式提速 1.21x–1.85x,其中 TerminalBench 上达到最高的 1.85x。\n\n## 轨迹处理器:把分支树压成「有界训练样本」\n\n弹性调度解决 GPU 利用率,轨迹处理器处理训练样本本身的爆炸。QwenGyre 在 proxy 这一层用 token-in\u002Ftoken-out(TITO)把 token ID 与 policy 的行为 log-prob 完整记录下来,把每个执行的结果存成「前缀共享的轨迹树」。\n\n有意思的是,超时或者评估器异常都没被当成「零分惩罚」。超时但工作区还在的 rollout,会以部分被保留的 workspace 作为评估对象,如果评估器自身崩溃,这条数据会被剔除,绝不污染策略梯度。每条执行最多留 5 条轨迹,优先级 main > main-summary > sub-agent > sub-agent-summary,共享目标只计一次,只有策略真正生成过的 token 参与训练损失。每个执行内部做损失平均,避免一条执行被扩张成多份以后权重被悄悄放大。\n\n这套约束有一个出乎意料的结论:把样本上限定为 5,实际能拿到的训练分数跟无上限的版本「对齐」,但前向-反向时间直接砍掉 25.2%。\n\n## 跑在 Qwen 3.8 2.4T 上的真实收益\n\n最能说明问题的还是 2.4T 模型本身。48 个训练步下来,QwenGyre 用 75.42 小时跑完,Async 需要 134.55 小时,Colocate 需要 91.47 小时,在 NL2RepoBench 上任务通过率从 52.48% 提到 58.54%——一个百分点不到 6 小时,这种边际成本,在 2.4T 的 MoE 上完全不可想象。训练用的还是改写过的 GSPO 目标,每组 16 个 rollout。\n\n横向看,veRL、AReaL、微软 Agent Lightning 都在做「异步 RL + Agent 训练」这件事:QwenGyre 把独占的弹性 GPU 调度拉到 Agent 训练里、把分支轨迹主动裁剪到有界尺寸,在 2.4T 的尺度上跑出公开可比较的端到端数字。对需要把编码、运维类 Agent 训出「真能干数小时活」的团队来说,这不再只是一个工程优化,而是判断「能不能训」的分水岭。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.33848","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":19,"name":20,"slug":20,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",{"id":22,"name":23,"slug":23,"description":14,"color":14},"c187600e-804c-4697-b828-1e4330e0eb10","qwen",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"fc071799-5ce9-45c9-ad95-c6199eab3ecf","en","Alibaba Qwen Releases QwenGyre Elastic RL: 1.85x Speedup for XLong-Horizon Agent Training","Alibaba's Qwen team released QwenGyre, an elastic reinforcement-learning framework built for hour-long agent rollouts. It dynamically swaps GPUs between rollout and training roles, and converts branched agent trajectories into bounded training samples. On the Qwen 3.8 2.4T model, NL2RepoBench pass rate rose from 52.5% to 58.5%.","Coding agents tasked with real software-engineering workloads often run rollouts that consume hours, hundreds of model and tool turns, and close to 1M tokens. This \"XLong-horizon\" regime is becoming the central bottleneck of agentic RL. GPUs sit idle waiting for the slowest rollout, or pile up empty bubbles of compute. Meanwhile, agents that mimic Claude Code compress context, spawn sub-agents and retry mid-run, so a single execution turns into a branching tree of shared-prefix trajectories, which explodes memory and compute if fully expanded. Alibaba's Qwen team just released QwenGyre to address both problems head-on.\n\n## Elastic scheduling: hardware follows the work\n\nQwenGyre splits multi-node GPUs into cells that share a training layout. Each cell can either serve rollouts on SGLang or run a full training step on Megatron. A core cell owns the authoritative weights and optimizer state; satellite cells serve rollouts by default and switch into the training role as soon as the scheduler judges that the remaining cells still cover all live requests.\n\nThe switch does not require the agent to cooperate. Every model call is routed through a proxy, and the proxy migrates KV-cache blocks over RDMA so prefix reuse stays high. Mooncake's Transfer Engine pulls core weights into a satellite so it can join an ongoing batch, and streaming training lets cells claim micro-steps as data lands. Average role switches took 8.52 seconds (rollout to training) and 3.46 seconds (training to rollout) in the team's runs.\n\nEnd-to-end speedups came in cleanly. Under matched GPU budgets and staleness, three benchmarks ran on Qwen 3.6 122B and Qwen 3.8 2.4T models: 1.38x–1.78x over an Async setup, and 1.21x–1.85x over a Colocate setup, with TerminalBench hitting the top 1.85x.\n\n## Trajectory processor: bounded samples out of a branching tree\n\nElastic scheduling fixes GPU utilization; the trajectory processor fixes training-sample explosion. At the proxy layer, QwenGyre uses token-in\u002Ftoken-out (TITO) to capture token IDs and the policy's behavior log-probs, and stores each execution as a prefix-shared trajectory tree.\n\nNotably, timeouts and evaluator errors are not punished as zero reward. A rollout that times out but whose workspace survives is still scored on the preserved state, and an evaluator crash causes the sample to be dropped rather than poisoning the gradient. Each execution keeps at most 5 trajectories, ranked main > main-summary > sub-agent > sub-agent-summary; shared targets are deduplicated, and only tokens the policy actually emitted contribute to the loss. Losses are averaged within each execution so expanding a tree does not silently inflate its weight.\n\nConstraining the cap to 5 preserved matched training scores from the uncapped run while cutting forward-backward time by 25.2%.\n\n## Real results on Qwen 3.8 2.4T\n\nThe 2.4T run is the most telling. Over 48 training steps, QwenGyre finished in 75.42 hours versus 134.55 hours for Async and 91.47 hours for Colocate, while lifting NL2RepoBench pass rate from 52.48% to 58.54% — about 5.4 hours per percentage point, marginal cost that is hard to imagine on a 2.4T MoE. Training used an adapted GSPO objective with 16 rollouts per group.\n\nSet against veRL, AReaL and Microsoft's Agent Lightning, QwenGyre is the framework that pulls exclusive elastic GPU scheduling into agent training and aggressively bounds branched trajectories. It is also the only one publishing comparable end-to-end numbers at the 2.4T scale. For teams trying to train agents that can really carry hours of coding or ops work, the question is no longer whether this is an optimization — it is whether the model can be trained at all.","alibaba-qwen-qwengyre-elastic-rl-xlong-agents","2026-10-14T00:00:00Z","2026-10-14T05:04:44.698290Z","2026-10-14T05:04:44.698304Z",true,"agent",52,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"206ea36a-1eea-463f-a241-1e3b32f5ec2d","USTC GraphForge:证据图把任务和 rubric 钉在一起,Qwen3.6-27B 涨三基准","graphforge-ustc-qwen36-27b-evidence-graph","2026-10-04T03:05:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"0d8fdf45-4585-47c0-9e78-3652e318b156","Apple Intelligence 中国版落地:通义千问接管语言 AI,百度负责视觉搜索","apple-intelligence-china-qwen-baidu-2026","2026-08-25T12:00:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"1844afb1-3a1c-4acd-9e4c-f5e2792a2018","下载免费不等于商用免费：HF Summer 2026 隐藏的开源前沿许可证分水岭","frontier-license-shift-hf-summer-2026","2026-08-23T12:30:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"4bb31ede-b9c4-4762-86ae-9d3b008557ca","Hugging Face Summer 2026 报告:Qwen 拿下 15 万衍生模型, GGUF 仓库一年涨 464%","hugging-face-state-of-open-models-summer-2026","2026-08-18T02:00:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"a2cd999e-e74f-43a9-a24b-3898cf5e9582","「0.8B 干到 16.72% WER」:Fun-ASR-Nano 用「端到端+RAG」把工业 ASR 卷出新尺度","qwen-fun-asr-nano-0-8b","2026-07-06T14:01:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"5edbfca7-7c1a-4df3-8577-a0195408b827","RSI-Anything:世界模型陪跑,自我改进有了回滚键","rsi-anything-world-model-recursive-self-improvement","2026-10-12T21:20:00+00:00"]