[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-21dcb43c-b3a5-4c22-9d89-fbf2bc283269":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},"21dcb43c-b3a5-4c22-9d89-fbf2bc283269","Agents-A1 把 35B MoE 推到 1T 级长程任务：上海 AI 实验室「水平扩展」路线开源","上海人工智能实验室 Agents-A1 团队（arXiv 2606.30616）给出反共识答案：一个 35B MoE 在 SEAL-0、IFBench、HiPhO、FrontierScience-Olympiad、MolBench-Bind 五个长程基准上领先 Kimi-K2.6、DeepSeek-V4-pro 等 1T 旗舰。\n\n方法是「Agent 水平扩展」：把轨迹拉长到平均 45K token，统一 6 个异构领域。训练三阶段——全领域 SFT 对齐 → 域级教师独立训练 → 多教师域路由 On-Policy 蒸馏，把 6 域装进同一 35B 学生。权重与管线全开源（HF: huggingface.co\u002FInternScience\u002Fagents-a1）。\n\n评论：LLM 进入 Agent 时代，性能瓶颈从「参数总量」转向「单次任务能调动的有效推理长度」。35B 用 45K token 视野下的多教师域路由蒸馏跑赢 1T，是「水平扩展」首次正面挑战「参数扩展」。后 Scaling Law 时代，Agent horizon 正在取代参数 scale 成为新主战场。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.30616","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",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},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source","2026-06-29T16:00:00Z","2026-06-30T16:17:58.383264Z","2026-06-30T16:17:58.383274Z",true,"agent",2]