8月25日 周二22 条
- Apple 推出搭载 M5 Max 与 M5 Ultra 的全新 Mac Studio
Apple 发布搭载 M5 Max 与全新 M5 Ultra 的 Mac Studio，AI 性能最高提升 4.3 倍，图形性能提升 1.8 倍，存储速度提升 2 倍。M5 Ultra 版本支持最高 512GB 统一内存与 1.2TB/s 内存带宽，可完全在设备端运行大型 LLM；四台集群可带来最高 3 倍分布式 AI 推理速度提升。新品今日起接受预购，9 月 22 日开售。 🔗 阅读原文 via AIHOT · https://aihot.virxact.com/items/cmt8qee503kvnro73qp7543yv
AI HOT 精选原文 ↗ - 基于阿里千问，路透社推出首个自研大模型
8月25日，据外媒报道，国际知名新闻机构汤森路透（路透社母公司）近日推出了其自研大模型Thomson-1。该模型基于阿里Qwen大模型开发，旨在降低对Claude等闭源模型的依赖并减少长期使用成本。 此前，汤森路透曾推出自己的AI助手，接入的是海外大模型厂商Anthropic的Claude大模型。出于降低闭源模型的依赖和成本因素考量，近期，汤森路透选择基于阿里开源模型Qwen3.5自研出Thomson-1 大模型。Thomson-1 将首先用于表格分析、文档审核等专业工作，并逐步承接部分此前由 Claude 完成的任务。 开源模型允许企业自行部署和调整，性价比高，安全可控。随着阿里Qwen等开源模型能力的持续增强，越来越多的企业和机构由闭源转向开源。美国互联网公司爱彼迎Airbnb CEO布莱恩·切斯基称公司正“大量依赖阿里巴巴的千问模型”，并直言“比OpenAI更好更便宜”；“美国版小红书”Pinterest也于近期转投Qwen，其CEO比尔·雷迪直言使用Qwen的成本不到Anthropic、OpenAI等同类闭源模型的8%。 截至目前，Qwen已经开源460多个模型，在开源社区总下…
雷峰网原文 ↗ - BIMSA 王雅晴：Scaling Law 触及天花板，「数据高效学习」指向 AI 的下一站｜IJCAI 2026
真正的智能，在于能否基于有限的数据，实现可靠的泛化与迁移。 作者丨张璐 编辑丨岑峰 幸丽娟 高质量标注数据快用完了，算力贵得离谱，黑盒依然无解。 大模型时代信奉的 Scaling Law，正在面临多重瓶颈。 更麻烦的是另一件事：很多任务天生就没有大数据——药物研发、冷启动推荐、科学探索。这些不是技术的短板，是世界的常态。 在IJCAI 2026大会的 Early Career Spotlight（早期职业焦点）环节，北京雁栖湖应用数学研究院（BIMSA）副教授王雅晴，给这些问题做出了系统回应。 王雅晴长期从事小样本学习与 AI for Science 研究。她博士毕业于香港科技大学，曾任百度研究院资深研究员，现任北京雁栖湖应用数学研究院（BIMSA）副教授。贯穿其研究经历的一条主线是：如何提高 AI 对有限数据与经验的利用效率，使其在监督稀缺的场景下仍能实现可靠的泛化、适应与决策。 她在题为《迈向高效数据智能：从少样本学习到智能体系统》（Toward Data-Efficient Intelligence: From Few-Shot Learning to Agentic Syste…
- 突破具身智能物理死角：IJCAI 2026 具身智能泛化与控制突破
告别理想仿真，中国学术团队用物理工程方案回应落地挑战。 作者丨张璐 编辑丨岑峰 过去一年多里，具身智能似乎按下了快进键。 从 Physical Intelligence 发布的π₀到开源社区追捧的OpenVLA，VLA大模型一路高歌猛进。在各种标准基准测试和精心搭建的实验室 Demo 中，机械臂叠衣服、拿取物品流畅得宛如人类，动辄刷出 90% 甚至 95% 以上的成功率。 然而，当全行业都在欢呼“机器人的 GPT 时刻”到来时，前沿学术界与产业界却爆发了一场剧烈的反思：这些动辄数十亿参数的具身大模型，真的学会“举一反三”了吗？还是在“死记硬背”？ 在主流学术视角中，VLA 大模型像是给机器人装上了具备常识与规划能力的“大脑”。但要让机械体真正干好精细活，仅有一个聪明的大脑远远不够。 最新实验数据表明，面对物品位置的随机挪动或新指令组合时，纯端到端大模型往往会受到“空间过拟合”的约束，末端执行成功率会出现剧烈衰减。 这恰恰说明：具身智能要真正走出实验室，不仅需要大模型提供常识和高层决策，更需要扎实的工程技术去打造精准、稳健的“小脑”。 下面雷峰网盘点入选 IJCAI 2026 的 6 项…
- Grafana 正式发布 gcx 和 MCP 服务器，助力基于遥测的智能代理开发
Grafana Labs 宣布"正式上线两款工具 gcx CLI 和 Grafana MCP 服务器，使 AI 编码代理能够在开发过程中实时查询可观测性数据。这两款工具均允许智能代理从 Grafana Cloud" 或自建技术栈中提取指标、日志、追踪、SLO 以及合成监控结果。这种集成有助于弥合智能代理生成代码的速度与工程师对代码功能理解之间的差距。据 InfoQ 报道"，这两款工具已于 4 月在 GrafanaCon 2026 大会上发布。 这些工具解决了基于智能代理的编码工作流中经常存在的一个问题。开发人员主要借助大语言模型（LLM）来规划、实现和审查变更。然而，它们可能在尚未完全理解代码变更的情况下就合并了拉取请求（PR）。过去，工程师在编写代码时会同步构建针对这次变更的心理模型。当代码差异由智能代理生成时，这一构建心理模型的过程便不复存在，尽管代码审查代理和“LGTM”审批会带来一种虚假的安全感。Grafana Labs 提出的解决方案是将可观测性数据纳入审查流程，作为一种额外的证据，用于对智能代理的输出结果进行检查，而非仅依赖代码审查。 从架构上讲，这两款工具的功能固化程度有…
InfoQ 中文原文 ↗
- 《有头有脸的大模型公司，集体搞起「匿名公测」》
头图来源：oxalpha.com 最近，海外开发者圈又被一个没有名字的模型刷屏了。 8 月 20 日，OpenRouter 上多了一个叫 Ox Alpha 的模型，100 万 token 上下文、支持图片和视频输入、能调用工具、免费。随后，它被接进 OpenCode、Hermes 等编程 Agent。有人拿它修代码，有人用它跑软件工程基准，还有人盯着 tokenizer 和报错信息，试图猜出它到底是谁家的。 目前，Ox Alpha 的身份还没有得到官方确认。社区最流行的猜测是，它与智谱 GLM-5.x 系列关系很近；依据是 tokenizer、推理模式报错和部分输出习惯的相似性。但这还只是技术分析，智谱官方目前还没有认领。 从 Ox Alpha 爆火，可以看到一个越来越常见的出海打法：模型先匿名上线，让海外开发者替它做第一轮测试、传播和身份鉴定。 一、从 HappyHorse 到 Ox Alpha Ox Alpha 并不是第一个蒙着脸跑出来的模型。 往前翻四个月，阿里的 HappyHorse（欢乐马）也是这个路数—— 没有发布会，没有公司名，直接空降到 AI 评测平台 Artifici…
极客公园原文 ↗
- ForeTime-VLA: Causal Future-Token Distillation from a World Action Model for Conveyor-Belt Manipulation
arXiv:2608.20735v2 Announce Type: replace-cross Abstract: Manipulating moving objects requires a policy to anticipate contact events, yet vision-language-action (VLA) policies are commonly fine-tuned from the current observation alone. World action models (WAMs) learn predictive dynamics, but running a video-scale teacher or explicitly imagining future frames at deployment is costly. We introduce ForeTime-VLA, a dense pi0.5 policy that distills a future-aware, action-equivalent representation fr…
arXiv cs.RO原文 ↗ - FlatVPR: Plug-and-play Geo-linear Residual Adapter for Geometric Rectification of Foundation Model Feature Manifolds
arXiv:2606.01734v2 Announce Type: replace-cross Abstract: This paper proposes ``FlatVPR,'' a novel geometric rectification paradigm that effectively bridges the trade-off between map lightweightness and localization accuracy in visual place recognition (VPR) by enforcing a feature manifold structure where any descriptor between two adjacent anchors $\mathbf{z}_A$ and $\mathbf{z}_B$ can be accurately reconstructed via linear interpolation $\hat{\mathbf{z}}_{pseudo} = (1-t)\mathbf{z}_A + t\mathbf{…
arXiv cs.RO原文 ↗ - PEAfowl: Perception-Enhanced Multi-View Vision-Language-Action for Bimanual Manipulation
arXiv:2601.17885v2 Announce Type: replace-cross Abstract: Bimanual manipulation in cluttered scenes requires policies that remain stable under occlusions, viewpoint changes and scene variations. Existing vision-language-action models often lack such robustness because (i) multi-view features are fused via view-agnostic token concatenation, yielding limited cross-view spatial representations, and (ii) language is injected as global conditioning, resulting in coarse instruction grounding. In this …
arXiv cs.RO原文 ↗ - GuideFetch: A Task Coordination Framework for Concurrent Navigation and Object Retrieval in Assistive Robot Dogs
arXiv:2608.18292v2 Announce Type: replace Abstract: Consider one robot guide dog escorting a blind user to a seat while a second retrieves and delivers an object. We introduce \textsc{GuideFetch}, a framework for coordinating this concurrent guide-and-fetch mission with heterogeneous robots. A large language model (LLM) instantiates a schedule-conditioned four-action schema; deterministic normalization and validation enforce registered targets, robot capabilities, and the selected schedule, whil…
arXiv cs.RO原文 ↗ - Latency-Tolerant Cloud-Edge Collaborative Vision-Language-Action Models via Emergent Representational Specialization
arXiv:2608.00569v2 Announce Type: replace Abstract: Deploying billion-parameter Vision-Language-Action (VLA) policies on mobile robots creates a systems conflict: semantic reasoning benefits from cloud GPUs, whereas closed-loop control must respond locally despite network delay and jitter. Existing hierarchical and asynchronous policies improve throughput, but their slow-path representations can still arrive stale or require explicit scheduling and delay cues. We introduce CloudEdgeVLA, a cloud-…
arXiv cs.RO原文 ↗ - Guided Action Flow: Q-Guided Inference for Flow-Matching Vision-Language-Action Policies
arXiv:2607.02092v3 Announce Type: replace Abstract: Deploying a pretrained flow-matching vision-language-action (VLA) policy on a particular robot and workspace often calls for task-specific adaptation, while full- policy fine-tuning is costly and changes the base behavior. We present Guided Action Flow, an inference-time method that keeps a pretrained SmolVLA policy frozen and steers its reverse-time action-flow sampling with gradients from a task-specific action-chunk critic. QGF trains a visu…
arXiv cs.RO原文 ↗ - Q-VGM: Q-Value-Gradient Matching for Offline-to-Online Reinforcement Learning of Flow-Matching VLA
arXiv:2606.08015v3 Announce Type: replace Abstract: We propose Q-Guided Value-Gradient Matching (Q-VGM), an offline-to-online reinforcement learning (RL) method for fine-tuning flow-matching vision-language-action (VLA) policies with a learned Q-function. Classical off-policy actor-critic methods improve a policy by following the critic gradient $\nabla_A Q$, but applying this update to flow policies requires backpropagation through the multi-step denoising process (BPTT), which is costly and un…
arXiv cs.RO原文 ↗ - Foundation Models in Robotics: A Comprehensive Review of Methods, Models, Datasets, Challenges and Future Research Directions
arXiv:2604.15395v3 Announce Type: replace Abstract: Over the recent years, the field of robotics has been undergoing a transformative paradigm shift from fixed, single-task, domain-specific solutions towards adaptive, multi-function, generalpurpose agents, capable of operating in complex, open-world, and dynamic environments. This tremendous advancement is primarily driven by the emergence of Foundation Models (FMs), i.e., large-scale neural-network architectures trained on massive, heterogeneou…
arXiv cs.RO原文 ↗ - When Should a Robot Think? Resource-Aware Reasoning via Reinforcement Learning for Embodied Robotic Decision-Making
arXiv:2603.16673v5 Announce Type: replace Abstract: Embodied robotic systems increasingly rely on large language model (LLM)-based agents to support high-level reasoning, planning, and decision-making during interactions with the environment. However, invoking LLM reasoning introduces substantial computational latency and resource overhead, which can interrupt action execution and reduce system reliability. Excessive reasoning may delay actions, while insufficient reasoning often leads to incorr…
arXiv cs.RO原文 ↗ - LaGEA: Language Guided Embodied Agents for Robotic Manipulation
arXiv:2509.23155v3 Announce Type: replace Abstract: Robotic manipulation benefits from foundation models that describe goals, but today's agents still lack a principled way to learn from their own mistakes. We ask whether natural language can serve as feedback, an error-reasoning signal that helps embodied agents diagnose what went wrong and correct course. We introduce LaGEA (Language Guided Embodied Agents), a framework that turns episodic, schema-constrained reflections from a vision language…
arXiv cs.RO原文 ↗ - LLM-Based Selection of Incongruent Verbal and Nonverbal Behavior for Virtual Humans
arXiv:2608.22731v1 Announce Type: cross Abstract: Nonverbal behavior generation systems for virtual agents often take an utterance as input and generate nonverbal behaviors that emphasize or illustrate the content of the verbal channel. However, human nonverbal behavior is shaped by more than the content of the speech. It is also influenced by speaker roles, interpersonal relationships, social context, and the cognitive and emotional states of the interactants. As a result, the nonverbal channel…
arXiv cs.RO原文 ↗ - ExploreAI: Agentic Exploration Knowledge Bases for Reproducible Observable-Regression Testing of Black-Box VR and 3D Applications
arXiv:2608.21628v1 Announce Type: cross Abstract: Black-box VR and 3D applications are difficult to regression test because observable failures depend on where a tester moves, what objects are visible, and which views are captured. Manual exploratory testing can find such failures, but its evidence is time-consuming to reproduce; systematic sweeps are reproducible, but they lack semantic guidance and spend exploration budget on low-value viewpoints. We observe that an LLM can make the high-level…
arXiv cs.RO原文 ↗ - Act with Intent: Distilling Behavior Intent for Vision-Language-Action Models
arXiv:2608.23478v1 Announce Type: new Abstract: Vision-Language-Action (VLA) models can turn multimodal context into robot actions, but their action decoders are still trained largely by behavior cloning. This supervises which motor command was demonstrated while leaving implicit the local objective served by the behavior under the instruction. Future-based supervision enriches action learning with frames, latent observations, trajectories, or motion representations, but these signals capture pa…
arXiv cs.RO原文 ↗