[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-edgewise-edge-agents-cloud-delegation":3,"topics-all":38,"news-related-97ee186c-a42b-4389-8427-3342d77e8590":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},"97ee186c-a42b-4389-8427-3342d77e8590","小模型何时喊云端大模型?NVIDIA 砍 62% 推理成本","NVIDIA 提出 EdgeWise 边云协作框架:Qwen3.5-9B 本地当主力,关键步骤才求助云端大模型。论文分析 2.4k 条编码 Agent 轨迹后用步级 RL 训 9B 模型自学何时委派,SWE-Rebench 推理成本最多降 62%,平均 46%,性能与全程用大模型相当。","长程编码 Agent 有个尴尬的成本结构:一个任务动辄几十上百步,业内惯常做法是把每一步都丢给云端前沿大模型。NVIDIA 的研究团队最近在 arXiv 上挂出论文 EdgeWise,开头就戳破这层窗户纸——不是每一步都需要大模型的能力,但每一步都在按大模型的价格付费。\n\n## 云端大模型的价值,集中在少数关键步骤\n\n论文先做了一个观察实验:分析 2.4k 条编码 Agent 轨迹,把每一步归入六类动作——Write\u002FEdit、Read\u002FSearch、Diagnose\u002FReproduce、Test\u002FValidate、Discover\u002FNavigate、Wait\u002FPoll,再测量每类步骤上云的价值。结论是:云端大模型的收益集中在少数关键步骤,云端调用率只要 40% 左右、把对的步骤交上去,整体表现就能与全程使用云端大模型相当,推理成本至少省 25%。\n\n## 让 9B 小模型自己学会何时求助\n\n基于这个观察,EdgeWise 把 Qwen3.5-9B 部署在本地当主力,仅在关键步骤把控制权交给云端大模型。怎么判断关键?论文给了两条路:一是外挂裁判,用轻量 critic 模型 GPT-5.4-mini 在线判断当前步骤是否需要更强模型;二是自委派,用步级 RL 把解题与委派一起训进 9B 模型本身,推理时不再需要额外决策模型。训练用 SWE-Smith 环境、Kimi-K3 当云端教师,2K 样本先 SFT,再在 40 张 A100 80GB 上跑 175 步 RL。消融显示:只做 SFT 的版本 pass@1 为 57.30%,加上 RL 后到 71.91%,成本只多花 3.74 美元——委派策略才是 RL 阶段真正的增量。\n\n## 数字:最多砍 62%,端侧部署再省一半\n\n主结果里,EdgeWise 在 SWE-Rebench 上最多降 62% 推理成本,Terminal-Bench 2.1 上降 34%,跨模型跨基准平均降 46%,性能与纯大模型执行持平或更好;云端模型换成 Claude Opus 4.8 的对照也列入主表。团队还把 Qwen3.5-9B 部署到一台 DGX Spark(vLLM 0.19.1、BF16、显存利用率 0.65)上做真实边缘验证,Pass@3 与 A100 环境相当,成本比全程云端再降约 50%。泛化性同样有数据:把本地模型换成 Tmax-9B,二选一委派把 pass@1 从 28.09% 拉到 74.16%,追平 GPT-5.5 的 74.16%,费用省 27.67 美元;裁判换成开源的 Qwen3.8-27B,效果也保持可比。\n\n## 所以呢\n\n这篇论文真正有意思的地方,不是又一个「小模型省钱」的故事,而是把「要不要求助」本身变成可训练的技能。对做 Agent 的人来说,接口账单的大头从来不是难题,而是被大材小用的常规步骤;对 NVIDIA 来说,这套框架天然给 DGX Spark 这类端侧设备铺路——本地 9B 常驻、云端按需升级,可能才是长程 Agent 的默认形态。论文:https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.17560","https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.17560","7437aeb9-930c-4866-a2e9-48003c1a792b",[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},"0a93ec8e-ea39-4693-81de-563ca8c173f7","inference",{"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},"25a4508f-7d83-4b9e-b899-6346614446a2","en","NVIDIA EdgeWise: SLMs Learn When to Delegate to Cloud","EdgeWise runs Qwen3.5-9B locally, delegating only key steps to a cloud LLM. Inference cost falls 62% on SWE-Rebench, 34% on Terminal-Bench 2.1.","Long-horizon coding agents have an awkward cost structure: a single task easily runs dozens or hundreds of steps, and the common practice is to send every step to a frontier cloud model. An NVIDIA research team recently posted a paper, EdgeWise, that calls out the problem in its opening — not every step needs frontier-level capability, yet every step is billed at frontier prices.\n\n## Cloud Value Concentrates in a Few Key Steps\n\nThe paper starts with an observation study: the authors analyze 2.4k coding-agent trajectories and classify each step into six action types — Write\u002FEdit, Read\u002FSearch, Diagnose\u002FReproduce, Test\u002FValidate, Discover\u002FNavigate, and Wait\u002FPoll — then measure the value of sending each type to the cloud. Their conclusion: the benefit of the cloud model concentrates in a small subset of critical steps. With a cloud usage rate of only around 40%, delegating the right steps yields performance comparable to using the cloud model for the entire trajectory, while cutting inference cost by at least 25%.\n\n## Teaching a 9B Model When to Ask for Help\n\nBased on that observation, EdgeWise deploys Qwen3.5-9B locally as the primary solver and hands control to a cloud LLM only at critical steps. How to decide which steps are critical? The paper offers two mechanisms. The first is an external judge: a lightweight critic model (GPT-5.4-mini) decides online whether the current step needs a stronger model. The second is self-delegation: a step-level RL method trains task solving and delegation jointly into the 9B model itself, so no extra decision model is needed at inference time. Training uses the SWE-Smith environment with Kimi-K3 as the cloud teacher: 2K examples for SFT, followed by 175 RL steps on 40 NVIDIA A100 80GB GPUs. The ablation is telling — the SFT-only model reaches 57.30% pass@1, while adding RL lifts it to 71.91% for just $3.74 more. The delegation policy, not the task skill, is the real delta of the RL stage.\n\n## The Numbers: Up to 62% Cheaper, and Half Again on the Edge\n\nIn the main results, EdgeWise reduces inference cost by up to 62% on SWE-Rebench and 34% on Terminal-Bench 2.1, averaging 46% across models and benchmarks, while matching or outperforming pure-LLM execution; a comparison with Claude Opus 4.8 as the cloud model is also included in the main table. The team further deploys Qwen3.5-9B on a single NVIDIA DGX Spark (vLLM 0.19.1, BF16, GPU memory utilization 0.65) for a real edge validation: Pass@3 stays comparable to the A100-based setup while cost drops roughly another 50% versus all-cloud execution. Generalization holds too — swapping the local model to Tmax-9B, binary delegation lifts pass@1 from 28.09% to 74.16%, tying GPT-5.5 at 74.16% while saving $27.67; replacing the judge with the open-source Qwen3.8-27B keeps performance comparable.\n\n## So What\n\nThe interesting part of this paper is not yet another \"small models are cheaper\" story — it turns \"should I ask for help\" itself into a trainable skill. For agent builders, the bulk of the API bill was never the hard problems; it was routine steps burning frontier-model tokens. For NVIDIA, the framework paves a natural road for devices like DGX Spark: a resident local 9B with on-demand cloud escalation may become the default shape of long-horizon agents. Paper: https:\u002F\u002Farxiv.org\u002Fabs\u002F2610.17560","edgewise-edge-agents-cloud-delegation","2026-10-16T21:07:57Z","2026-10-16T21:08:01.159847Z","2026-10-16T21:08:01.159860Z",true,"agent",157,[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},"74cb8016-bdf7-4c8b-adbc-456d8c899b1a","EdgeWise 把 9B 模型训练成「喊人」高手:边云协同砍半成本","nvidia-edgewise-edge-cloud-slm-self-delegation-coding","2026-10-19T01:01:45+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"ed39ed38-b5fa-4f58-92cf-d05233ab998b","Speculate with Memory：LLM Agent 无损加速 2.5×，准确率涨 39pp","speculate-with-memory-llm-agent-acceleration","2026-07-15T08:15:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"addfda39-d64f-4cb9-809a-7dfadd72d421","ECHO:百度把多Token预测搬进主干最后一层","echo-baidu-multi-token-prediction-self-recursion","2026-10-21T01:15:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"ab1e6bdf-9433-41f0-be4d-cc0f78523ac2","SonicBoom v0.0.3:一个 .so 干掉手工调优","sonicboom-v003-mlir-compiler","2026-10-19T09:03:05+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"9a77e0d7-8164-494e-984d-54cb57a7a0dd","Magnitude 开源:本地 Agent 专用推理引擎","magnitude-self-tuning-local-agent-inference-engine","2026-10-01T17:08:35+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"5fa17d26-6090-4805-893a-aa880cf33369","让 Agent 学会遗忘:删掉旧推理,分数反而涨了","agents-forget-reasoning-iclr-compression","2026-09-28T19:30:00+00:00"]