Long-context inference and Retrieval-Augmented Generation (RAG) handle evidence selection at vastly different scales, fr…

Perplexity Research and turbopuffer have released **pplx-embed-v2-context-9b-preview**, a contextual embedding model for…

While Large Language Models (LLMs) advertise million-token context windows, reasoning quality often collapses as inputs …

Training a Mixture-of-Experts (MoE) model at long context or large batch size fails as soon as any one component's peak …

Recently, linear attention layers have been increasingly adopted to replace softmax attention at scale for long-context …

Few-shot in-context learning (ICL), the capacity of a model to infer abstract patterns from input-output examples provid…

Reasoning vision-language models (VLMs) process images and videos as long sequences of visual tokens, making inference e…

StepFun has released Step 5 Preview, a sparse Mixture-of-Experts model with 600B total parameters and 27B active per tok…

Alibaba's Qwen3.8-Omni-Flash understands audio and video, plans tasks, calls tools, and reports about 45.7% fewer tokens…

Enabling robots to adapt to unfamiliar environments as readily as humans remains a moonshot goal of embodied AI. No fini…

Scaling large language models efficiently has motivated sparse capacity mechanisms such as Mixture-of-Experts and, more …

Post-training attention sparsification reduces the quadratic cumulative attention cost of pretrained Transformers by sel…

Speculative decoding accelerates rollout generation, which dominates the cost of reinforcement learning (RL) post-traini…

Text-driven 3D generation has advanced rapidly in creating large-scale outdoor environments and detailed indoor scenes, …

发布不到两年，MCP 从 Anthropic 的一个开放标准长成了 AI 应用连接层的事实标准——又在 2026-07-28 的规范里把自己重构了一遍：会话没了、握手没了、sampling 和 roots 被废弃。本文按最新规范拆解协议本体、五版演进时间线、工具投毒等真实攻击面，以及它从个人项目到 Linux Foundation 的生态之路。

《Language Models are Few-Shot Learners》（Brown et al., 2020）把 GPT-3 推到 175B 参数，真正的发现却是另一个：只靠提示词里放几个例子，模型就能完成没训过的任务——in-context learning 从此改写了 NLP 的工作方式。本文精读这篇论文的机制、数据与遗留争议。

Linear attention enables efficient long-context autoregressive decoding by compressing history into recurrent states, bu…

Agent 工程系统学习第 5 章：工具是 Agent 的手脚。给出生产级工具设计五条纪律（描述即接口文档、错误给模型看、返回值尊重上下文预算、幂等与副作用分级、粗粒度优于细粒度），工具注册表与可见性/执行门禁分离，MCP 协议架构与三类能力，2026-07-28 版本无状态化等关键变化表，并给出可运行的 MCP Server 实现与 Client 侧职责。

Language model agents are deployed with a harness, the software around the model that manages its context, tools, and fe…

Streaming world models should render a place consistently across repeated visits. Directly supervising such revisits req…

An assistant that serves the same user over a long horizon has to answer from what that user has revealed: which prefere…

Pretrained transformers use little of their depth to follow references in context. Thirteen base models reliably follow …

Coding agents now run for hours, not minutes. Every edit, test run and log read goes back into the model’s context. A te…

系列第 ③ 篇，对应技能地图第 2 格「Grounding」。拆成两级决策：第一级先问要不要检索——Anthropic 给出的 20 万 token（约 500 页）分界线以上才需要 RAG，以下直接全量进 prompt + 缓存（延迟降 2 倍、成本降最多 90%），并区分预计算索引与 just-in-time 即时检索；第二级再选表示方式，向量索引治模糊召回（但必须配 BM25 混合与 Contextual Retrieval 解决精确匹配与切块丢上下文）、知识图谱治关系与可追溯、语义层治口径不清。附可量化收益表（检索失败率 5.7% → 3.7% → 2.9% → 1.9%）、四个实现注意项、context rot 与上下文压缩/笔记/子智能体三件套，以及一张可抄的选型决策树。

Interactive video world models must maintain broad scene context under camera motion while producing high-fidelity obser…

Agent harnesses (the system prompt, tool set, execution hooks, and context-management scaffolding around a model) are a …

Linear attention is increasingly used in frontier language models for efficient long-context inference and constant-memo…

Can one language model hand its live memory to another without the receiver rereading the context? We demonstrate useful…

Sequential memory agents process long documents by reading chunks one after another while maintaining a compact memory s…

Many NLP tasks, such as summarization and extractive question answering, reduce to retrieving relevant content from docu…

Anthropic 发布模型上下文协议（Model Context Protocol），以开放标准统一 LLM 应用与外部数据源、工具的连接方式，被社区称为'AI 应用的 USB-C 接口'。

Back in June, Nvidia announced that it had secured agreements with Microsoft and a host of other PC makers to develop AI…

At a glance Harnessed Agentic RL: Microsoft Research Asia introduces a training paradigm in which the same agent harness…

!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…

推理模型改写了提示词的规则——OpenAI 官方文档明确写着「避免思维链提示」「通常不需要 few-shot 示例」；Anthropic 则把叙事重心移交给了「上下文工程」。提示工程死了吗？没有，它收缩了：从话术技巧收缩成「先建评估、再写规格、管好上下文」的工程实践。本文以三家官方文档为一手依据，对照出仍然有效的技巧清单、官方明确说可以省掉的技巧，以及实证研究给出的边界。

KV Cache 让一次推理不用重算历史 token，但多轮对话每次都把 10 万字的 system prompt 重发一遍——前缀缓存回答的是「相同前缀算一遍就够，能不能跨请求共享」。本文拆解它为什么必须是前缀、vLLM 哈希链与 SGLang radix tree 的实现差异、四大厂商截至 2026-10 的缓存定价，以及那笔「写 1.25 倍、读 0.1 倍」的账什么时候是赚的、什么时候反亏 25%。

在提示里放几个输入-输出示例，模型一次前向传播就把新任务干得像模像样——没有梯度更新，没有训练循环，这就是上下文学习（ICL）。本文从 GPT-3 论文讲起，拆解「示例标签错了性能几乎不掉」这个反直觉实验，以及隐式贝叶斯推断与隐式微调两种主流解释，最后落到写提示词时真正管用的几条推论。

!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…

!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…

TechCrunch Disrupt 2026** is seven days away, bringing 10,000+ founders, investors, operators, and tech leaders to Mosco…

AI 改不动大项目的首要原因不是模型不够聪明，而是它看不到相关的代码。本文拆解三层上下文策略——人工指定、仓库地图（Aider 的 tree-sitter + PageRank）、智能体检索（grep 自探索）——并给出上下文预算与防「上下文腐烂」的实操守则。

Mamba 把 SSM 参数改成输入的函数，让固定大小的状态学会按内容取舍；再靠并行扫描与 kernel 融合把状态装进 SRAM，线性复杂度落地——3B 匹敌两倍大的 Transformer，5 倍推理吞吐，线性扩展到百万长度。文末梳理截至 2026-10 它与 Attention 的分工现状。

从自注意力不识顺序的痛点讲起，用二维复数把旋转位置编码的推导一次讲透：内积为何只依赖 m−n、高维频率如何分配，配一份本地跑通的 numpy 最小实现与整体平移不变性实验，再谈 PI、NTK-aware、YaRN 的长上下文扩展路线与 ALiBi 的取舍。

从 Peebles 与谢赛宁的 DiT 论文到 Sora 的时空 patch，讲清 Diffusion Transformer 的三个关键设计：patch 化、adaLN-Zero 条件注入与以计算量为标尺的可扩展性。

零样本语音克隆只需约 6 秒参考音频：说话人编码器把音色压成一个向量，再去条件化 TTS 生成——XTTS 一条路是「嵌入式克隆」，OpenVoice 一条路是「合成后换色」。本文拆解两条技术路线的原理、效果边界，以及这道技术必须面对的滥用与合规问题。

MCP 连接智能体与工具，A2A 负责智能体之间怎么对话。Google 发起、现由 Linux Foundation 旗下 AAIF 治理，最新版本 1.0（截至 2026-10）。Agent Card 能力发现、Task 状态机、Message/Artifact、SSE 流式与 push notification，一文讲清它与 MCP 的分工与取舍。

Claude Code 用法实操：CLAUDE.md 记忆文件分层加载规则与模板、子智能体定义与上下文隔离、PreToolUse/PostToolUse 钩子拦截与自动格式化配置，附探索-计划-实现-提交循环、headless CI 集成与成本边界分析。

!(https://www.gstatic.com/images/branding/googleg/1x/googleg_standard_color_128dp.png)Add as a preferredsource on Google…

Sensors are useful not only for understanding the world but also for deciding what to do next. Existing sensor models ho…

A central capability of embodied agents is to accomplish complex objectives through sequences of interdependent tasks. Y…