[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-nvidia-molt-agentic-rl-pytorch":3,"news-related-4bb93e6c-a626-4e8e-b5f3-3ea3a15c8604":36},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"4bb93e6c-a626-4e8e-b5f3-3ea3a15c8604","NVIDIA Molt 把 Agentic RL 训练栈打回 PyTorch 原生:一个异步循环训多模态 MoE,代码小到 LLM 一次读完","NVIDIA 把 Agentic RL 的\"训练-部署\"接口打回到 PyTorch 原生层,今天和 Hugging Face 一起放出了 Molt 框架。arXiv 2607.21653,7 月 22 日提交,GitHub 仓库 NVIDIA-NeMo\u002Flabs-molt 已开放。\n\n传统 Agent RL 框架是 trainer \u002F 分布式后端 \u002F rollout 三层叠加,每改一次算法都得串通整条流水线。Molt 反过来:整个 agent 就是个普通程序,一个异步循环就同时跑多模态与 MoE 策略,token、policy 版本、模型语义三件套在 rollout 与训练之间严格一致——保证模型绝不拿自己没生成过的 token 去更新权重。\n\n实现上,Molt 不绑 Megatron 也不绑 vLLM 任何特定栈,而是\"代码够小、够干净\",让 AI 编程助手能一次性读完整个仓库。在 matched 异步协议下,Molt 与基于 Megatron 的 SOTA 栈在统计意义上持平,leanness 几乎不付出性能代价。开源协议 + 现成 recipes + 容器一并提供,Agent RL 研究者不必再为分布式胶水代码发愁。\n\n值得注意的两点。第一,这是 NVIDIA 第一次把\"agent 本身是一段普通程序\"作为一等公民的框架,意味着 agent loop 里的工具调用、外部环境交互、长 horizon 决策可以无缝接进 RL 训练流,而不是外挂一层适配器。第二,async rollout + MoE + 多模态的组合,在 NeMo AutoModel 之后进一步把 NVIDIA 的训练栈推向\"全异步、原生 PyTorch\"的统一方向,大模型 RLHF\u002FAgent 训练开始摆脱对专用 RL 框架的依赖。\n\n落地价值:Molt 直接戳中\"Agent RL 复现难、研究迭代贵\"的老毛病。当框架小到能被 LLM 一次性理解,Agent RL 才能从少数大公司的工程壁垒变成普通实验室可重复的实验——这是 2026 下半年 Agent 走向规模化研究的关键基建。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.21653","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},"8dac812d-3839-4abe-a855-5f56ec9515fd","nvidia",{"id":21,"name":22,"slug":22,"description":13,"color":13},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":28},"7da196e0-b054-42c9-a51d-b784644ad9ce","en","NVIDIA Molt rebuilds agentic RL on native PyTorch, minimal code","NVIDIA has reset the \"training–deployment\" interface of Agentic RL to the native PyTorch layer, and today released the Molt framework together with Hugging Face. arXiv 2607.21653, submitted July 22; the GitHub repo NVIDIA-NeMo\u002Flabs-molt is now public. Traditional Agent RL frameworks stack trainer \u002F distributed backend \u002F rollout in three layers, so every algorithmic change has to ripple through the entire pipeline. Molt inverts this: the whole agent is just an ordinary program, and a single async loop simultaneously drives multimodal and MoE policies. Token, policy version, and model semantics remain strictly consistent between rollout and training — guaranteeing the model never updates weights against tokens it didn't generate. Implementation-wise, Molt isn't tied to Megatron or vLLM — instead it's \"small enough and clean enough\" for an AI coding assistant to read the entire repo in one go. Under a matched async protocol, Molt is statistically on par with a Megatron-based SOTA stack, paying virtually no performance cost for its leanness. An open-source license, ready-to-use recipes, and containers are all provided — so Agent RL researchers no longer have to wrestle with distributed glue code. Two points worth flagging. First, this is the first time NVIDIA has made \"an agent is just an ordinary program\" a first-class citizen in a framework, meaning tool calls, external environment interaction, and long-horizon decisions inside the agent loop can flow seamlessly into RL training, rather than being wrapped by an external adapter. Second, the combination of async rollout + MoE + multimodal pushes NVIDIA's training stack further into the \"fully async, native PyTorch\" unified direction after NeMo AutoModel, freeing large-model RLHF\u002FAgent training from its dependence on specialized RL frameworks. Bottom line: Molt hits the long-standing pain point of \"Agent RL is hard to reproduce and expensive to iterate on\". When a framework becomes small enough to be understood by an LLM in one read, Agent RL stops being an engineering moat for a few big companies and becomes a reproducible experiment for ordinary labs — a key piece of infrastructure as Agent research scales in the second half of 2026.","nvidia-molt-agentic-rl-pytorch","2026-07-28T03:00:00Z","2026-07-28T00:05:51.166843Z","2026-08-19T02:08:40.142862Z",true,"agent",204,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"f26ace13-9c96-47ea-a528-b6682a22aa1e","Apodex 1.1 把推理搬进真实执行:PIVOT-RL 定位关键决策点,35B mini 开源","apodex-1-1-agentic-execution-pivot-rl","2026-08-25T14:30:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"32b938b6-01a3-43c9-b040-14db6c5f57c6","NVIDIA 把 Agent 装进一个 Python 类:被忽略的 NOOA,一半 token 跑出 SWE-bench 82.2%","nvidia-nooa-python-agent-framework","2026-08-23T17:20:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"ce70384a-990b-4994-bfb6-27775be45661","TensorRT Edge-LLM 0.10.0：边端第一个统一的 C++ 多模态推理栈","tensorrt-edge-llm-0-10-multimodal-runtime","2026-08-23T00:00:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"a8b9d045-0f4c-4596-baa7-060955365877","TensorRT Edge-LLM 0.10.0：Qwen3.8-27B Day-0 上车，边缘 LLM 推理再加速","tensorrt-edge-llm-qwen3-8-27b-day0","2026-08-21T15:00:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"e8965513-b56f-475b-b15f-22a5ea2d2a4e","Agent 取代人成为 HF Hub 一号用户:Claude Code 占 44.4%,还有一次 4.5 天未察觉的入侵","hf-hub-agent-user-claude-code-4-5-day-intrusion","2026-08-21T08:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"b4754043-6b19-499f-8459-f8fc786f4d80","Pokee-Isaac 28B 把 10M 上下文塞进客户边界:28B 参数在 RULER 10M 上 93.3%","pokee-isaac-28b-10m-context","2026-08-20T14:00:00+00:00"]