Agent swarms and the new model economics

這是一篇基於 Cursor 部落格研究內容的技術分析。

AI語音進入“表演時代”：阿里Qwen-Audio-3.0-TTS登頂全球權威榜單

📌 【阿里巴巴 Qwen-Audio】從「能說話」進化到「會表演」，Qwen-Audio-3.0-TTS 登頂權威榜單

资讯Qwen

DiFA: Inference-Time Forward-Process Alignment for Diffusion Models

📌 【DiFA 研究】無需重新訓練，透過前向過程對齊提升擴散模型生成品質

browser-use/video-use

📌 【開源專案】透過 Claude Code 進行影片剪輯：video-use 讓 AI 直接接管 ffmpeg

Python

Distilled Reinforcement Learning for LLM Post-training

📌 【研究論文】結合教師指導與 RL：DistilledRL 解決 LLM 蒸餾的知識轉移困境

Environment-free Synthetic Data Generation for API-Calling Agents

📌 【新研究】無需執行環境，利用 LLM 模擬數位世界來生成 API 呼叫資料

EvolvingWorld: An Open-Schema Framework for Co-Evolving Role-Play Agents and World Model in Interactive Literary World

📌 【EvolvingWorld】突破靜態模擬：實現角色與世界觀的長期共演

Introducing Cosmos 3 Edge

這篇內容屬於開源專案（GitHub/Hugging Face）與產業新聞的結合，重點在於 NVIDIA 發布的新模型及其在邊緣運算（Edge Computing）的應用能力。

Guides

這份素材屬於「部落格文章／指南」，將聚焦於 Simon Willison 分享的編碼代理（Coding Agents）實踐模式。

JoyNexus: Service-Oriented Multi-Tenant Post-Training for VLA Models

📌 【研究論文】JoyNexus：透過多租戶架構解決 VLA 模型訓練的資源效率問題

microsoft/AI-Engineering-Coach

📌 【Microsoft 開源專案】AI Engineer Coach：打造你的 AI 程式碼助手專屬儀錶板

TypeScript

NVIDIA Releases Cosmos 3 Edge: A 4B-Parameter Open World Model That Reasons and Generates Robot Actions On-Device

這篇內容屬於「產業新聞」，我將針對 NVIDIA 發布 Cosmos 3 Edge 模型及其技術架構進行轉寫。

AI InfrastructureLanguage ModelLarge Language ModelPhysical AI

LVSum: A Benchmark for Timestamp-Aware Long Video Summarization

📌 【Apple ML 最新研究】LVSum：長影片摘要的挑戰，不在於看懂畫面，而在於對齊時間點

PrefectHQ/fastmcp

內容型別判斷：開源專案 (GitHub)

Python

ReflectWorld-MM: An Entity-Oriented Multimodal Memory System for Open-Ended Video Streams

📌 【新論文】ReflectWorld-MM：以實體為核心，讓 AI 助手能持續觀察並理解世界

QwenLM/qwen-code

📌 【Alibaba Qwen 開源新專案】打造住在 Terminal 裡的 AI 編碼代理 (Coding Agent)

TypeScript

Safety and alignment in an era of long-horizon models

📌 【OpenAI 研究分享】面對長期目標模型 (Long-horizon Models)，現有的安全對齊機制還夠用嗎？

Safety

TimeLens2: Generalist Video Temporal Grounding with Multimodal LLMs

📌 【TimeLens2】打破 MLLM 侷限：讓多模態大模型能精準定位影片中的時間區間

vercel-labs/deepsec

📌 【Vercel Labs 開源專案】DeepSec：用 Agent 驅動的大規模程式碼漏洞掃描工具

TypeScript

Length Value Model: Scalable Value Pretraining for Token-Level Length Modeling

📌 【Apple ML 研究】LenVM：將生成長度轉化為價值預估，實現 Token 等級的精準控制

10 Open-Source No-Code AI Platforms for Building LLM Apps, RAG Systems, and AI Agents

📌 【開源盤點】10 款免寫程式碼的開源 AI 平臺，涵蓋 LLM 應用、RAG 與 Agent

Low Code AINo Code AITopVibe Coding

AstrBotDevs/AstrBot

📌 【開源專案】AstrBot：一個串接多平臺的開源 Agent 聊天機器人框架

Python

bojieli/ai-agent-book

📌 【開源專案】一本全開源的 AI Agent 實戰書：從公式到程式碼

Python

Canner/WrenAI

📌 【Canner 開源】WrenAI：讓 AI Agent 生成並治理 BI 儀錶板的 GenBI 引擎

Python

Cura 1T: Specialized Model for Agentic Healthcare

📌 【HuggingFace Daily Papers】Cura 1T：用自進化迴圈打造醫療專用 Agentic LLM

From Human-Centric to Agentic Code Review: The Impact of Different Generations of Generative AI Technology on Review Quality

📌 【研究論文】AI 程式碼審查變快了，但審查品質真的變好了嗎？

jamiepine/voicebox

📌 【開源專案】Voicebox：本地端跑完整的 AI 語音輸入輸出堆疊

TypeScript

Kimi K3 vs DeepSeek V4 Pro vs GLM-5.2: Open Trillion-Scale MoE Models Compared on Benchmarks, License, and Serving Cost

📌 【MarkTechPost 整理】Kimi K3、DeepSeek V4 Pro、GLM-5.2 三款兆級 MoE 開源模型實測比較

Agentic AILanguage ModelLarge Language Model

DSWorld: A Data Science World Model for Efficient Autonomous Agents

📌 【HuggingFace Daily Papers】DSWorld：用世界模型替資料科學代理省下大量運算

kvcache-ai/ktransformers

📌 【kvcache-ai 開源】KTransformers：用 CPU-GPU 異質運算跑 LLM 推理與微調

Python

On-Policy Delta Distillation

📌 【HuggingFace Papers】用 Delta 訊號重做 On-Policy Distillation，推理 LLM 短期後訓練更強

Perplexity AI Releases WANDR: An Open Benchmark Evaluating Research Agents That Must Search Wide And Deep

📌 【Perplexity AI 發布】WANDR：評測研究代理「廣而深」搜尋能力的開放基準

Agentic AI

Q00/ouroboros

📌 【Q00 開源】ouroboros：用規格優先取代即興提示的 Agent OS

Python

Loop the Loopies!

📌 【HuggingFace Daily Papers】Loopie：迄今最強的迴圈 Transformer 登場

Recursive Harness Self-Improvement

📌 【HuggingFace Daily Papers】用 RHI 讓輕量 Agent 超越高推理成本設定

S1-Omni: A Unified Multimodal Reasoning Model for Scientific Understanding, Prediction, and Generation

📌 【HuggingFace Papers】S1-Omni：統一多模態科學推理的單一模型

Someone Fine-Tuned OpenBMB’s MiniCPM5-1B on Claude Fable 5 Traces to Ship a 657MB Local Thinking Model

📌 【社群專案】用 Claude 輸出微調 MiniCPM5-1B，推出 657MB 本地思考模型

Language ModelLarge Language ModelMachine LearningSoftware Engineering

topoteretes/cognee

📌 【topoteretes 開源】Cognee：讓 AI 代理擁有跨會話的長期記憶

Python

Xiaomi-Robotics-1: Scaling Vision-Language-Action Models with over 100K Hours of Real-World Trajectories

📌 【Xiaomi】百萬小時真實軌跡訓練的 VLA 機器人基礎模型

Understanding Reasoning from Pretraining to Post-Training

📌 【HuggingFace Papers】用西洋棋當實驗場，看清 Pretraining 如何決定 RL 成效

anthropics/skills

📌 【Anthropic 開源】anthropics/skills：Claude 專用技能資料夾實作範例

Python

agentscope-ai/agentscope

📌 【agentscope-ai】AgentScope 2.0：為生產環境打造的 Agent 框架

Python

Controlling Reasoning Effort in LLMs

📌 【Sebastian Raschka 專欄】如何打造有多檔推理力度的 LLM

DLLM-Searcher: Adapting Diffusion Language Model for Efficient Search Agents

📌 【Semantic Scholar 論文】用擴散語言模型打造更高效的搜尋代理

elder-plinius/G0DM0D3

📌 【開源專案】G0DM0D3：把多模型聊天介面推向紅隊與認知研究

TypeScript

Hierarchical Denoising For Multi-Step Visual Reasoning

📌 【HuggingFace Papers】分層去噪讓影片模型學會多步視覺推理

Kimi: Threat or menace?

📌 【TechCrunch 觀察】Kimi K3 開源釋出，華爾街與政界為何緊張？

Government & Policydavid sacksdean ballkimi

Mayor Mamdani Says Landlords Can't Use AI Images to Advertise

📌 【紐約市新政】房東用 AI 修圖招租須強制揭露

KnockOutEZ/wigolo

📌 【開源專案】wigolo：讓 AI 代理擁有本地優先的網頁操作層

TypeScript

MoonshotAI/kimi-cli

📌 【MoonshotAI 開源】Kimi CLI：終端機裡的 AI 程式開發代理

Python

musistudio/claude-code-router

📌 【開源專案】claude-code-router：把多個編碼 Agent 接到本地統一路由層

TypeScript

ogulcancelik/herdr

📌 【開源專案】herdr：終端機裡的 agent 多工管理器

Rust

rivet-dev/agentos

📌 【rivet-dev 開源】agentOS：跑在程式內的 AI 代理作業系統

Rust

Setting up your spare Mac for Claude Code to control, a step-by-step guide

📌 【實戰指南】把閒置 Mac 交給 Claude Code 全權控管

SearchOS-V1: Towards Robust Open-Domain Information-Seeking Agent Collaboration

📌 【HuggingFace Daily Papers】SearchOS-V1：把搜尋進度變成共享狀態，解決代理卡死迴圈

SigNoz/signoz

📌 【SigNoz】用 OpenTelemetry 打造一體化開源可觀測性平臺

TypeScript

upstash/context7

📌 【upstash 開源】Context7：把最新程式庫檔案直接餵進 LLM 提示

TypeScript

WAIC 2026“鎮館之寶”STEPX Neo亮相，引領人機互動新範式

📌 【WAIC 2026】STEPX Neo 獲鎮館之寶，L3 級原生 AI 手機亮相

资讯阶跃星辰

warpdotdev/warp

📌 【warpdotdev】Warp 開源：從終端機長出的 agentic 開發環境

Rust

A scorecard for the AI age

📌 【OpenAI 部落格】AI 時代的實用 ROI 評分卡

Company

AsySplat: Efficient Asymmetric 3D Gaussian Splatting for Long-Sequence Scene Modeling

📌 【HuggingFace Papers】AsySplat：非對稱 3D Gaussian Splatting 砍掉多餘運算

Beyond Success Rate: Cost-Aware Evaluation of Offensive and Defensive Security Agents

📌 【arXiv 研究】安全代理評估不能只看成功率，還要算成本

cs.CRcs.AI

Build an Agentic Event Venue Operator with MongoDB Atlas, Voyage, and LangGraph

📌 【MarkTechPost 教學】用 MongoDB、Voyage、LangGraph 打造有記憶的活動場館 Agent

Agentic AIAI AgentsLanguage ModelPromote

How Smartsheet built a remote MCP server on AWS

📌 【AWS ML 案例】Smartsheet 在 AWS 上打造遠端 MCP 伺服器

Advanced (300)Amazon BedrockAWS FargateBest Practices

Chat2Scenic: An Iterative RAG-Based Framework for Scenario Generation in Autonomous Driving

📌 【HuggingFace Papers】Chat2Scenic：用迭代 RAG 框架生成自駕測試場景指令碼

LongStraw: Long-Context RL Beyond 2M Tokens under a Fixed GPU Budget

📌 【HuggingFace Papers】固定 GPU 預算下，長上下文 RL 突破 2M tokens

Fine-tune video and image models at scale with NVIDIA NeMo Automodel and 🤗 Diffusers

📌 【NVIDIA × Hugging Face】用 NeMo Automodel 大規模微調 Diffusers 擴散模型

n8n-io/n8n

📌 【開源專案】n8n：用視覺畫布與自寫程式碼打造 AI Agent 工作流

TypeScript

NVIDIA AI Releases Nemotron 3 Embed: An Open Embedding Collection Whose 8B Checkpoint Ranks #1 on RTEB

📌 【NVIDIA 釋出 Nemotron 3 Embed】8B 嵌入模型登頂 RTEB 檢索榜

Embedding ModelLanguage ModelLarge Language ModelMachine Learning

On Locality and Length Generalization in Visual Reasoning

📌 【視覺推理研究】區域性感知如何讓模型真正學會舉一反三？

rohitg00/ai-engineering-from-scratch

📌 【GitHub 熱門】從零手刻 AI 工程：503 堂課、20 階段免費開源課綱

Python

Sakana AI’s Error Diffusion Trains Dale-Compliant Dual-Stream Networks, Reaching 96.7% MNIST and 61.7% CIFAR-10 Without Backpropagation

📌 【Sakana AI】不用反向傳播，雙流網路在 MNIST 達 96.7%

AI InfrastructureAI Paper SummaryLanguage ModelLarge Language Model

RxBrain: Embodied Cognition Foundation Model with Joint Language-Visual Reasoning and Imagination

📌 【HuggingFace Papers】RxBrain：用語言與視覺想像統一體現認知規劃

Show Me Examples: Inferring Visual Concepts from Image Sets

📌 【Apple ML】從圖片集合中推論視覺概念：VICIS 新任務與架構

SUFLECA: Scaling Up Feature Learning for CAD-to-image Alignment

📌 【HuggingFace Papers】SUFLECA：用幾何特徵擴大學習規模，零射 CAD 對齊超越全監督

Token Time Continuous Diffusion for Language Modeling

📌 【HuggingFace Papers】Token Time Continuous Diffusion：用連續空間與逐 token 時間提速擴散語言模型

Zyphra Releases ZUNA1.1: An Apache 2.0 EEG Foundation Model With Variable-Length Inputs From 0.5 To 30 Seconds

📌 【Zyphra 開源】ZUNA1.1：支援 0.5–30 秒可變長度輸入的 Apache 2.0 EEG 基礎模型

Language ModelLarge Language ModelMachine LearningOpen Source

tirth8205/code-review-graph

📌 【開源專案】code-review-graph：用結構地圖讓 AI 審 code 不再狂燒 token

Python

給AI發工號、定崗位、做績效，數字員工終於能落地了

📌 【面壁智慧開源】給 AI 發工號定崗位，StaffDeck 讓數字員工落地

资讯AgentStaffDeck清华大学

[AINews] Kimi K3 2.8T-A50B: the largest open model ever released; Opus 4.8-class at Sonnet 5 pricing

📌 【Moonshot AI】Kimi K3 2.8T-A50B：號稱史上最大開放權重模型

browseros-ai/BrowserOS

📌 【開源專案】BrowserOS：讓 AI 代理操作你已登入的瀏覽器帳號

TypeScript

Embarrassingly Simple Self-Distillation Improves Code Generation

📌 【Apple ML 研究】只用自身輸出微調，LLM 程式碼生成大幅進步

GRASP: GRanularity-Aware Search Policy for Agentic RAG

📌 【HuggingFace Daily Papers】GRASP：用 RL 訓練 Agent 自適應協調檢索粒度

Introducing Grok on Amazon Bedrock

📌 【AWS ML】xAI Grok 4.3 登陸 Amazon Bedrock

Amazon BedrockAnnouncementsIntermediate (200)

How Cars24 scales conversations and builds faster with OpenAI

📌 【OpenAI 案例】Cars24 用語音與聊天 Agent 月處理百萬分鐘對話

memvid/memvid

📌 【memvid 開源】把 AI 記憶打包成單一檔案，免資料庫又能即時檢索

Rust

Moonshot AI Releases Kimi K3: A 2.8 Trillion Parameter Open MoE Model With Kimi Delta Attention and 1M Context

📌 【Moonshot AI 釋出 Kimi K3】2.8 兆引數開源 MoE，首見 3T 級模型

Agentic AILanguage ModelLarge Language ModelMachine Learning

NousResearch/hermes-agent

📌 【NousResearch 開源】Hermes Agent：會自己長出技能的 AI 代理

Python

NVIDIA Nemotron 3 Embed Ranks #1 Overall on RTEB, Advancing Agentic Retrieval

📌 【NVIDIA】Nemotron 3 Embed 登頂 RTEB，主攻 Agentic 檢索

OpenAI Details GPT-Red: An Internal Automated Red-Teaming Model That Beat Human Red-Teamers 84% To 13% On Prompt Injection

📌 【OpenAI 內部揭露】GPT-Red 自動紅隊模型在 Prompt Injection 上以 84% 擊敗人類 13%

Agentic AILanguage ModelMachine LearningSecurity

Scaling Agentic AI Factories Through Extreme Co-Design with NVIDIA BlueField

📌 【NVIDIA 開發者部落格】用 BlueField 極致協同設計擴充套件 Agentic AI 工廠

Agentic AI / Generative AIData Center / CloudNetworking / CommunicationsAI Agent

SEED: Self-Evolving On-Policy Distillation for Agentic Reinforcement Learning

📌 【HuggingFace Papers】SEED：用後驗技能蒸餾補齊 Agent RL 監督缺口

Smarter and Cheaper at Once: Byte-Exact KV-Cache Grafting Turns a Frozen Small Model into a Verified-Knowledge Flywheel

📌 【HuggingFace Daily Papers】凍結小模型不動權重，靠 KV-Cache 嫁接同時變強又變省

Spectral Rewiring for Exploration, Purification, and Model Merging

📌 【HuggingFace Papers】用 Spectral Rewiring 只留推理核心，免訓練強化多域能力

🔬 The Lab of the Future Should Feel Like a Data Center — Andy Beam & Rafa Gómez-Bombarelli, Lila Sciences

📌 【Latent Space 報導】未來實驗室長得像資料中心？Lila Sciences 的 AI 科學工廠

UniVR: Thinking in Visual Space for Unified Visual Reasoning

📌 【HuggingFace Daily Papers】UniVR：在純視覺空間中統合推理與規劃

VideoChat3: Fully Open Video MLLM for Efficient and Generalist Video Understanding

📌 【HuggingFace Papers】VideoChat3：4B 引數全開源影片 MLLM 兼顧效率與泛化

WanSong v1.0 Technical Report

📌 【HuggingFace Daily Papers】WanSong v1.0：純擴散模型單次生成 5 分鐘雙音軌歌曲

全球首個！銀河通用新框架僅需人類影片即可部署，特斯拉蚌埠住了

📌 【銀河通用】全球首個具身智慧測試時後訓練框架 WAM-TTT 釋出

资讯首页轮播具身智能银河通用

[AINews] Thinky's Inkling: 975B-A41B multimodal, new best American Apache 2.0 open model (with Inkling-Small, 276B-A12B)

📌 【Latent Space 報導】Thinky 釋出 Inkling：975B-A41B 美國 Apache 2.0 開放多模態模型

Boogu-Image-0.1: Boosting Open-Source Unified Multimodal Understanding and Generation

📌 【開源多模態】Boogu-Image-0.1：低運算預算逼近封閉系統的統一模型

CLaRa: Bridging Retrieval and Generation with Continuous Latent Reasoning

📌 【Apple ML 論文】CLaRa：用連續潛在推理打通檢索與生成

Google Releases LiteRT.js: A JavaScript Binding of LiteRT That Runs .tflite Models in Browsers via WebGPU

📌 【Google 發布】LiteRT.js：用 WebGPU 在瀏覽器跑 .tflite 模型

Software Engineering

GPT-Red: Unlocking Self-Improvement for Robustness

📌 【OpenAI 新專案】GPT-Red 用 Self-Play 自動紅隊提升安全性

Safety

datawhalechina/hello-agents

📌 【Datawhale 開源】從零動手打造 AI Native Agent 的系統教程

Python

HKUDS/nanobot

📌 【HKUDS 開源】nanobot：可自架、輕量的個人 AI 代理執行環境

Python

Inkling – Open-Weights 975B Parameter LLM

📌 【Thinking Machines】Inkling：開放權重 975B MoE 多模態模型

Model Routing Is Simple. Until It Isn’t.

📌 【HuggingFace Blog】模型路由看似簡單，實則是系統最佳化難題

One Layer Is Enough: Adapting Pretrained Visual Encoders for Image Generation

📌 【Apple ML】一層注意力就夠？用預訓練視覺編碼器做影像生成

ShortOPD: Recovering Pruned LLMs with Short-to-Long On-Policy Distillation

📌 【HuggingFace Daily Papers】剪枝後 LLM 崩潰？ShortOPD 用短到長蒸餾救回生成能力

snap-stanford/Biomni

📌 【Stanford SNAP】Biomni：可自主執行生醫研究的通用 AI Agent

Python

SpaceXAI Open-Sources Grok Build: The Rust Agent Harness, TUI, and Tool Layer Behind Its Coding CLI

📌 【SpaceXAI 開源】Grok Build：Rust 打造的 AI 編碼代理框架與工具層

Agentic AIAI AgentsMachine LearningOpen Source

Soofi Consortium Releases Soofi S 30B-A3B: An Open Hybrid Mamba-Transformer MoE Foundation Model For German And English

📌 【Soofi Consortium 開源】Soofi S 30B-A3B：德英雙語混合 Mamba-Transformer MoE 基礎模型

Agentic AIAI InfrastructureAI Paper SummaryLanguage Model

Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, Inkling

📌 【Thinking Machines 首款開放權重模型 Inkling】反對一刀切 AI 的 MoE 路線

TCMira Muratithinking machine labs

Tracing Agentic Failure from the Flow of Success

📌 【HuggingFace Daily Papers】只用成功軌跡，就能找出 Agent 在哪一步失敗？

Thinking Machines Lab Releases Inkling: A 975B-Parameter Open-Weights Multimodal MoE With 41B Active Parameters And Controllable Thinking Effort

📌 【Thinking Machines Lab 釋出 Inkling】975B 多模態 MoE 開放權重，思考力可調

Agentic AIAI InfrastructureLanguage ModelLarge Language Model

Uncertainty Quantification for LLM Function-Calling

📌 【Apple ML】LLM Function-Calling 的 Uncertainty Quantification 首評

vllm-project/vllm

📌 【vLLM】高吞吐 LLM 推論與服務的開源利器

Python

Welcome Inkling by Thinking Machines

📌 【HuggingFace 報導】Thinking Machines 開源 Inkling：近 1T 引數原生多模態模型

12 Ways to Reduce LLM Latency and Inference Costs in Production

📌 LLM 生產降本：先量測再砍浪費

Cursor 0day: When Full Disclosure Becomes the Only Protection Left

📌 Cursor 0day：開專案自動執行惡意 git

agentgateway/agentgateway

📌 agentgateway：Agentic AI 完整連線方案

Rust

earendil-works/pi

📌 Pi 代理框架：自擴充套件編碼代理

TypeScript

google/skills

📌 【Google 開源】google/skills 彙整雲端代理技能

Python

How to manage AI investments in the agentic era

📌 【OpenAI】代理時代如何管理 AI 投資

AI Adoption

Hy-Embodied-VLM-1.0: Efficient Physical-World Agents

📌 Hy-Embodied-VLM-1.0：3B 高效具身代理

cs.CV

How to Run an Autoresearch Workflow with RL Agent Skills and NVIDIA NeMo

📌 自主 Agent 自動化 RL 研究

Agentic AI / Generative AIDeveloper Tools & TechniquesAgent SkillAI Agent

kangarooking/cangjie-skill

📌 書與影片蒸餾成可呼叫 AI Skills

Python

Multi-agent social intelligence with Strands Agents and Amazon Bedrock

📌 多代理系統自動化社群情報

Customer SolutionsExpert (400)Partner solutionsStrands Agents

Lessons From the Leaderboard: What 5,000+ Kagglers Taught Us About Improving AI Reasoning

📌 【NVIDIA】5,000+ Kaggler 的推理工程實戰

Agentic AI / Generative AIAI AgentAI InferenceBlackwell

openclaw/openclaw

📌 OpenClaw：本地優先的個人 AI 助手

TypeScript

Multilingual Semantic Retrieval for Apple Music Search

📌 【Apple ML】混合檢索助 Apple Music 搜尋轉換增 2.28%

openinterpreter/openinterpreter

📌 Open Interpreter 轉 Rust，最佳化低成本模型

Rust

Proactive Agent Research Environment: Simulating Active Users to Evaluate Proactive Assistants

📌 【Apple ML】Pare：模擬主動使用者評估助理

Post-Train NVIDIA Cosmos 3 in One Day Using Agent Skills

📌 【NVIDIA】一天後訓練 Cosmos 3 準確率達 93%

Agentic AI / Generative AIComputer Vision / Video AnalyticsSimulation / Modeling / DesignAI Agent

Scaling medical content review at Flo Health with Amazon Bedrock – Part 2

📌 Bedrock 讓醫療內容審查快 60%

Amazon BedrockCustomer Solutions

Read It Back: Pretrained MLLMs Are Zero-Shot Reward Models for Text-to-Image Generation

📌 預訓練 MLLM 作零樣本獎勵模型

Scaling UX testing with Amazon Nova Act: A new approach to user flow analysis

📌 【Amazon】Nova Act 視覺導覽 UX 測試

Advanced (300)Amazon BedrockAmazon DynamoDBAmazon Nova

又快又聰明，阿里釋出Qwen-Audio-3.0-Realtime：即時語音大模型四項功能升級

我們需要根據提供的資訊撰寫繁體中文技術部落格文章。首先判斷內容型別：來源是「量子位」的一篇報導，標題是「又快又聰明，阿里釋出Qwen-Audio-3.0-Realtime：即時語音大模型四項功能升級」，內容是關於阿里巴巴釋出一個即時語音互動對話模型的新聞/部落格報導。所以屬於「產業新聞／部落格報導」…

资讯阿里巴巴

ABot-AgentOS: A General Robotic Agent OS with Lifelong Multi-modal Memory

📌 【新論文】ABot-AgentOS：透過多模態記憶與自演化機制，打造機器人通用的代理作業系統

Agent要數量也要腦子！浪潮資訊一邊單櫃養4萬Agent，一邊讓大模型組隊答題

📌 【浪潮資訊】Agent 時代來臨，AI 基礎設施正從 GPU 中心轉向多元算力協同

资讯浪潮信息

Anthropic Claude Sonnet 5 vs Sonnet 4.6 vs Opus 4.8: Agentic Coding Benchmarks, API Pricing, and Cost-Performance Tradeoffs Compared

📌 【Anthropic 新品發布】Sonnet 5 登場：更強的 Agentic 能力與成本效能權衡

Agentic AILanguage ModelLarge Language ModelSoftware Engineering

Beyond the Single Camera: Agentic Multi-View Reasoning in Sports Video Understanding

📌 多視角體育影片理解基準與 Agent

cs.CV

Building a VideoAgent-Style Multi-Agent System: Intent Parsing, Graph Planning, and Tool Routing for Video Editing Tasks

這篇素材屬於技術教學／開源專案實作說明（Tutorial/Implementation Guide），內容聚焦於如何建構一個模仿 VideoAgent 流程的多代理系統（Multi-Agent System），用於處理影片理解、檢索與編輯任務。

Agentic AIAI AgentsTutorials

Building an agentic AI solution at Bluesight with Amazon Bedrock

這是一篇針對「產業新聞／部落格報導」型別的技術文章。

Amazon BedrockAmazon Bedrock AgentCoreAmazon CloudWatchAmazon Cognito

Building and shipping Mac and iOS apps without opening Xcode

這篇內容屬於「產業新聞／部落格報導」，重點在於分享一種開發流程的實作方法。

Jul 13, 2026Societal ImpactsClaude’s values across models and languages

📌 【Anthropic】軸線量化 Claude 價值觀跨模型與語言

Research

Extreme Event Likelihoods with Guided Generative Models

📌 【NVIDIA】引導擴散估極端事件機率

Simulation / Modeling / DesignClimate / Weather / Ocean ModelingNGC

katanemo/plano

📌 Plano：Agent 生產化統一資料平面

Rust

Mistral AI Releases Robostral Navigate: An 8B Model Enabling Robots to Navigate Complex Environments Using a Single RGB Camera

這是一篇關於 Mistral AI 最新發布的機器人導航模型的技術報導。

Agentic AIPhysical AIRobotics

Read It Back: Pretrained MLLMs Are Zero-Shot Reward Models for Text-to-Image Generation

📌 【arXiv 研究】不用訓練！將多模態大模型直接轉化為影像生成 RL 的報酬模型

cs.CV

Prime Intellect Releases Verifiers v1: Composable Tasksets, Harnesses, and Runtimes for Agentic RL Training and Evaluations

📌 【Prime Intellect】Verifiers v1：解耦任務、Harness 與 Runtime

Agentic AISoftware Engineering

simonlin1212/TradingAgents-astock

根據您提供的資訊，這是一個開源專案（GitHub Project），我將依照開源專案的架構進行轉寫。

Python

NVIDIA Ising Decoding Cuts Color Code Logical Error Rates by Over 300X

📌 【NVIDIA】Ising 解碼器讓色碼邏輯錯誤率降 347 倍

Agentic AI / Generative AIData Center / CloudDeveloper Tools & TechniquesAI Foundation Models

Skyfall AI Releases MORPHEUS: A Persistent Enterprise Simulation Benchmark That Makes Continual Reinforcement Learning Necessary Under Structured Non-Stationarity

📌 【Skyfall AI 新發布】MORPHEUS 基準測試：解決 RL 在真實企業環境中的「非平穩性」挑戰

Agentic AIAI Paper Summary

Stanford Researchers Introduce TRACE: A Capability-Targeted Agentic Training System That Turns Recurrent Agent Failures Into Synthetic RL Environment

📌 【Stanford】TRACE：失敗驅動的 Agent 能力訓練

Agentic AIAI InfrastructureAI Paper SummaryLanguage Model

TauricResearch/TradingAgents

📌 【開源專案】TradingAgents：整合多種 LLM 與資料來源的金融交易多代理人框架

Python

Vexa-ai/vexa

這是一個開源專案（GitHub），我將依照開源專案的骨架進行撰寫。

Python

Weak-to-Strong Generalization via Direct On-Policy Distillation

📌 【Weak-to-Strong Generalization】利用 RL 政策位移，將小模型成果蒸餾至大模型

Claude Code sends 33k tokens before reading the prompt; OpenCode sends 7k

📌 Claude Code 與 OpenCode Token 消耗對比：前者單次請求多出 2.6 萬個 Token

A Sovereign, Open-Source Foundation Model for German and English

📌 SoofiS‑30B‑A3B：結合 MoE 與 Mamba 的開源德英雙語基礎模型

Comfy-Org/ComfyUI

📌 ComfyUI：視覺創作的高模組化節點圖引擎

Python

From RGB Generation to Dense Field Readout: Pixel-Space Dense Prediction with Text-to-Image Models

📌 從 RGB 生成到稠密讀出：利用文字生成模型直接預測畫素空間稠密任務

Dicklesworthstone/destructive_command_guard

📌 防止 AI 程式編寫助手誤刪檔案的「Destructive Command Guard」

Rust

google-gemini/gemini-cli

📌 Gemini CLI：在終端機直接呼叫 Google Gemini 大模型

TypeScript

Guides

📌 掌握 Coding Agents 的實務模式：從 Claude Code 到 OpenAI Codex 的高效開發指南

heygen-com/hyperframes

📌 HyperFrames：用 HTML+CSS 直接產出確定性 MP4，讓 AI 代理也能寫影片

TypeScript

HKUDS/Vibe-Trading

📌 Vibe‑Trading：一鍵式個人化交易代理，提供完整交易功能

Python

InsForge/InsForge

📌 InsForge：讓 AI Coding Agent 擁有完整後端操作能力的開源平臺

TypeScript

microsoft/mxc

📌 Microsoft eXecution Container (MXE) — 跨平臺沙盒執行框架的早期預覽

Rust

Migrating a production AI agent to GPT-5.6: 2.2x faster, 27% cheaper

📌 GPT‑5.6 Sol 讓 Ploy 代理上線速度提升 2.2 倍、成本下降 27%

KronQ: LLM Quantization via Kronecker-Factored Hessian

📌 KronQ：結合 Kronecker‑Hessian 與梯度共變異的 LLM 後訓練量化新框架

ruvnet/ruflo

📌 Ruflo：為 Claude Code 與 Codex 加上 100+ 專屬 Agent 的協調層與自學記憶

TypeScript

Self-Guided Test-Time Training for Long-Context LLMs

📌 Self‑Guided Test‑Time Training 讓長上下文 LLM 表現提升最高 15%

teng-lin/notebooklm-py

📌 【開源專案】notebooklm-py：用 Python 程式化操作 NotebookLM 的非官方 API

Python

Scalable Visual Pretraining for Language Intelligence

📌 挑戰純文字預訓練：將視覺檔案直接轉化為語言智慧的 scalable 路徑

virattt/ai-hedge-fund

📌 AI Hedge Fund：以多位投資大師為藍本的教學型交易決策系統

Python

Trust Region Policy Distillation

📌 解決 On-Policy Distillation 不穩定問題：Trust Region Policy Distillation (TOP-D)

“母病速歸”式 AI：號稱省 65% Token，實測只有 8.5%

📌 “母病速歸”式 AI：號稱省 65% Token，實測只有 8.5%

Ant Group’s Robbyant Unveils LingBot-VA 2.0: A Causal Video-Action Model Built Natively for Physical AI

📌 Ant Group Robbyant 發布 LingBot‑VA 2.0：首個原生因果影片‑動作模型，針對實體機器人操作打造

AI InfrastructureAI Paper SummaryPhysical AIPromote

ARDY: Autoregressive Diffusion with Hybrid Representation for Interactive Human Motion Generation

📌 ARDY：即時文字與動作限制控制的自回歸擴散式 3D 人體動作生成框架

ATH-MaaS/Pixelle-Video

📌 【Pixelle-Video】一鍵產出 AI 完全自動短影片的開源引擎

Python

DayuanJiang/next-ai-draw-io

📌 用自然語言畫圖：Next AI Draw.io 讓 AI 直接生成 draw.io 流程圖

TypeScript

Do Automated Evals Work?

📌 Do Automated Evals Work?：自動化評估到底能否可信？

Enhancing In-context Panoramic Generation via Geometric-aware Pretraining

📌 全景影像生成新突破：Canvas360 透過幾何感知預訓練提升一致性

Compete Then Collaborate: Frontier AI Teachers Build a Verifiable Curriculum to Improve a Coding Student Beyond Imitation

📌 競爭後合作：前沿 AI 教師打造可驗證課程，提升程式碼學生超越單純模仿

harry0703/MoneyPrinterTurbo

📌 MoneyPrinterTurbo：只給關鍵詞就自動產出高畫質短影片

Python

Ideas Have Genomes: Benchmarking Scientific Lineage Reasoning and Lineage-Grounded Idea Generation

📌 科學靈感也有基因？新基準測試挑戰 AI 的「科學譜系推理」能力

langflow-ai/langflow

📌 Langflow：視覺化打造 AI Agent 與工作流的全功能平臺

Python

LongE2V: Long-Horizon Event-based Video Reconstruction, Prediction, and Frame Interpolation with Video Diffusion Models

📌 LongE2V：利用影片 Diffusion 模型實現長時程事件相機影像重建與預測

OpenCoF: Learning to Reason Through Video Generation

📌 OpenCoF：透過影片生成提升時間推理能力

openai/codex

📌 OpenAI Codex CLI：在本機跑的程式碼助理

Rust

Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents

📌 解決「行為狀態衰減」：主動記憶代理讓長程任務不再遺忘

anthropics/claude-code

📌 Claude Code：終端機 AI 代理，讓程式碼編寫更快、更自然

Python

Shubhamsaboo/awesome-llm-apps

📌 不再從零開始：100+ 個可直接執行的 LLM 應用範本庫

Python

stablyai/orca

📌 【Stability AI 開源】Orca：讓多個 AI Agent 並行運作的協作編排器

TypeScript

SAM-MT: Real-Time Interactive Multi-Target Video Segmentation

📌 SAM-MT：打破目標數量限制，實現即時多目標影片分割

UniClawBench: A Universal Benchmark for Proactive Agents on Real-World Tasks

📌 UniClawBench：用即時 Docker 容器評測，衡量真實環境下的主動式 AI 代理人

volcengine/OpenViking

📌 【ByteDance 開源】OpenViking：為 AI Agent 量身打造的上下文資料庫

Python

A Quantized Native Runtime for On-Device Semantic Audio Generation

📌 端側部署新突破：原生量化執行時實現語音生成

aaif-goose/goose

📌 Goose：基於 Rust 的本機 AI Agent，整合 MCP 與 15+ 模型

Rust

agentscope-ai/agentscope

📌 AgentScope 2.0 釋出：面向生產環境的靈活 Agent 架構

Python

Accelerating End-to-End Co-Folding Performance with NVIDIA BioNeMo Agent Toolkit

📌 NVIDIA BioNeMo：讓 AI Agent 跑生物預測不再卡關

Agentic AI / Generative AIDeveloper Tools & TechniquesSimulation / Modeling / DesignBioNeMo

AI Model Co-Design: Hardware-Friendly LLM Design

📌 NVIDIA 硬體協同設計：Blackwell 高效能 LLM 關鍵策略

Agentic AI / Generative AIData Center / CloudDeveloper Tools & TechniquesAI Inference

[AINews] OpenAI launches GPT 5.6 Sol/Terra/Luna, Codex becomes ChatGPT superapp

📌 OpenAI 發布 GPT‑5.6 系列：Sol、Terra、Luna，Codex 變身 ChatGPT 超級應用

Build a semantic layer for agentic AI on AWS with Stardog and Amazon Bedrock AgentCore

📌 AWS 實戰：用 Stardog 與 AgentCore 打造 Agentic AI 語意層

Advanced (300)Amazon AthenaAmazon BedrockAmazon Bedrock AgentCore

ChromeDevTools/chrome-devtools-mcp

📌 ChromeDevTools 推出 MCP Server，讓 AI 直接控制瀏覽器

TypeScript

Disaggregated prefill and decode for LLM inference on SageMaker HyperPod

📌 SageMaker 實戰：用 DPD 解耦 LLM 推理延遲

Advanced (300)Amazon Elastic Kubernetes ServiceAmazon SageMaker HyperPodAmazon Simple Storage Service (S3)

CineMobile: On-Device Image-to-Video Diffusion for Cinematic Camera Motion Generation

📌 CineMobile：讓電影級鏡頭移動在手機上跑

DrugGen 2: A disease-aware language model for enhancing drug discovery

📌 DrugGen 2：結合疾病 Ontology 與 GRPO 強化藥物發現

Flash-BoN: Instant Drafts for Inference-Time Scaling in Diffusion Models

📌 Flash-BoN：用「草稿加速」讓擴散模型生成效率翻倍

jamiepine/voicebox

📌 jamiepine/Voicebox：本地全棧語音工作室，一次整合 7 款 TTS 引擎

TypeScript

Jet-Long: Efficient Long-Context Extension with Dynamic Bifocal RoPE

📌 Jet-Long：動態雙焦 RoPE 實現高效長上下文處理

microsoft/agent-governance-toolkit

📌 Microsoft Agent Governance Toolkit：讓 AI Agent 安全上線

Python

microsoft/graphrag

📌 Microsoft 推出 GraphRAG：用知識圖譜強化 LLM 推理能力

Python

NVIDIA/OpenShell

📌 NVIDIA OpenShell：為自主 AI Agent 提供安全沙箱執行環境

Rust

NVIDIA/skills

📌 NVIDIA 推出 Agent Skills：讓 AI 代理者精準掌握 CUDA 與平臺工具

Python

pytorch/pytorch

📌 PyTorch 核心架構解析：Tensor、Autograd 與 TorchScript 如何協作？

Python

wonderwhy-er/DesktopCommanderMCP

📌 Desktop Commander MCP：把 AI 編輯、終端指令與檔案管理全收進同一個聊天介面

TypeScript

Accelerating End-to-End Co-Folding Performance with NVIDIA BioNeMo Agent Toolkit

生物分子結構預測與共摺（co‑folding）正成為藥物發現與蛋白質設計的核心工作流程。為了讓 AI 代理人能夠高效執行完整的管線——包括多重序列比對（MSA）生成、共摺推理、服務以及多 GPU 規模擴充套件——每個環節都必須兼具速度與記憶體效率。NVIDIA 近期發布的 BioNeMo Agent…

Agentic AI / Generative AIDeveloper Tools & TechniquesSimulation / Modeling / DesignBioNeMo

AI Model Co-Design: Hardware-Friendly LLM Design

We need to produce a blog article in Traditional Chinese, based on the given info. Must follow the "撰寫規範" from system message, which includes: source…

Agentic AI / Generative AIData Center / CloudDeveloper Tools & TechniquesAI Inference

Datalab Lift vs the Field: How a 9B Schema-First Extractor Compares with NuExtract3, LlamaExtract, Marker, and Docling

📌 Datalab Lift 與其他工具比較：9B Schema‑First 抽取器如何簡化檔案處理

Large Language ModelMachine LearningOCR

CineMobile: On-Device Image-to-Video Diffusion for Cinematic Camera Motion Generation

CineMobile 論文提出一套在行動裝置上執行的影像到影像（I2V）擴散框架，透過「蒸餾導向修剪」、「擴散蒸餾」與「混合量化」三項技術，在保持視覺品質的同時顯著提升運算速度，使電影感鏡頭移動的影片生成成為行動裝置可行的應用。

DrugGen 2: A disease-aware language model for enhancing drug discovery

📌 DrugGen 2：結合疾病本體與蛋白序列的藥物分子生成模型

Disaggregated prefill and decode for LLM inference on SageMaker HyperPod

📌 【AWS 技術分享】拆分 Prefill 與 Decode，解決 LLM 推理的延遲尖峰問題

Advanced (300)Amazon Elastic Kubernetes ServiceAmazon SageMaker HyperPodAmazon Simple Storage Service (S3)

Flash-BoN: Instant Drafts for Inference-Time Scaling in Diffusion Models

📌 Flash‑BoN：用即時草稿加速擴散模型推理，效能突破固定時限

Incentivizing Temporal-Awareness in Egocentric Video Understanding Models

📌 【Apple 研究】解決 MLLM 「無視時序」問題：用 RLVR 強化第一視角影片理解

Jet-Long: Efficient Long-Context Extension with Dynamic Bifocal RoPE

📌 Jet-Long：透過動態雙焦 RoPE 實現高效的長文本擴充套件

jamiepine/voicebox

📌 本地化全棧語音工作室 Voicebox：整合 7 種 TTS 引擎，打造私有的語音 I/O 流程

TypeScript

MCP tool design: Practical approaches and tradeoffs

📌 【AWS 技術分享】MCP 工具設計：別讓 API 直接暴露給 LLM，小心 context 崩潰

Advanced (300)Best Practices

Meta Superintelligence Labs Releases Muse Spark 1.1: A Multimodal Reasoning Model for Agentic Tasks on Meta Model API

📌 Muse Spark 1.1 閉源推理模型

AI InfrastructureModel Context Protocol (MCP)Software EngineeringUncategorized

microsoft/graphrag

microsoft/graphrag

Python

Meet LingBot-World-Infinity: An Open Causal World Model With An Agentic Harness

📌 【螞蟻集團】LingBot-World-Infinity：用 MoBA Attention 解決世界模型「長程漂移」

Agentic AIAI InfrastructureAI Paper SummaryLanguage Model

NVIDIA/skills

📌 NVIDIA 發布官方 AI Agent Skills 目錄，提供即用的 CUDA‑X 與 AI Blueprint 教學

Python

Recursive Language Models Meet Uncertainty: The Surprising Effectiveness of Self-Reflective Program Search for Long Context

本文提出 Self‑Reflective Program Search for Long Context (SRLM)，透過自我一致性、推理追蹤長度與語言化信心三種內在不確定性訊號，對遞迴語言模型（RLM）的程式搜尋進行不確定性感知的引導。在相同時間預算下，SRLM 在多種長文本基準與不同骨幹模型上…

Reducing High-Bandwidth Memory Bottlenecks in JAX-Based LLM Training with Host Offloading

📌 JAX 主機解除安裝緩解 HBM 瓶頸，提升大模型訓練吞吐

Agentic AI / Generative AIDeveloper Tools & TechniquesSimulation / Modeling / DesignLLM Techniques

Robbyant Releases LingBot-VLA 2.0: An Open-Source 6B Vision-Language-Action (VLA) Model for Cross-Embodiment Robot Manipulation

📌 【Ant Group】開源 LingBot-VLA 2.0：6B 引數模型實現跨機體機器人操控

Agentic AIAI InfrastructureAI Paper SummaryEmbedding Model

Synthetic Data Generation for Financial AI Research with NVIDIA NeMo

📌 【NVIDIA 技術分享】解決金融資料不均：用 NeMo 打造 50 萬筆高品質合成資料集

Agentic AI / Generative AIData ScienceAI AgentAI-Ready Data

Unmasking On-Policy Distillation: Where It Helps, Where It Hurts, and Why

📌 Unmasking On-Policy Distillation：診斷框架揭示何時有幫助、何時有害

cline/cline

📌 跨平臺開源 Coding Agent：Cline 將 AI 助手整合至 IDE 與終端機

TypeScript

[AINews] Lilian Weng summarizes 35 papers on Harness Engineering for RSI

📌 Lilian Weng 彙整 35 篇論文：為什麼 RSI 的關鍵在於 Harness Engineering？

Create a LangChain Deep Agents Harness Profile for NVIDIA Nemotron 3 Ultra to Improve Performance

📌 【NVIDIA】不靠微調，用 Harness Profile 工程提升 Nemotron 3 Ultra 的 Agent 準確率

Agentic AI / Generative AITop StoriesAI AgentLangChain

jingyaogong/minimind

📌 MiniMind：只用 3 塊錢與 2 小時即可在個人 GPU 上訓練 64M 超小語言模型

Python

Data for Agents

📌 【NVIDIA 觀點】AI Agent 的核心不在權重，而是在於高品質的開放資料

Graphify-Labs/graphify

📌 別再用 Grep 找程式碼了：用 Graphify 將整個專案轉化為可查詢的知識圖譜

Python

Jul 8, 2026AlignmentAn off switch for dual use knowledge in AI models

📌 【Anthropic 研究】不只是拒絕，而是讓 AI 具備可開關的「雙用途知識」

Research

LMCache/LMCache

LMCache 是一個開源的 KV Cache 管理層，將原本臨時的 KV Cache 轉換為可持久化、可跨多個推理引擎重複使用的 AI‑native 知識。透過可觀測性堆疊與生成品質轉換，它能降低 Time‑To‑First‑Token（TTFT）並提升吞吐量，特別適用於長上下文代理、多輪對話及…

Python

microsoft/SkillOpt

📌 【Microsoft 開源】SkillOpt：像訓練神經網路一樣，讓 AI Agent 自我演化技能

Python

Native-speed vLLM transformers modeling backend

📌 【HuggingFace】vLLM transformers 後端效能追平，模型部署不再需要手動移植

NVIDIA/SkillSpector

📌 NVIDIA SkillSpector：AI Agent 技能安全掃描工具

Python

OpenAI releases new voice models for more natural live conversations

📌 OpenAI 發布 GPT-Live-1 語音模型，提升即時對話自然度

AppsChatGPTvoice model

Securing Amazon Bedrock AgentCore Runtime with AWS WAF

📌 【AWS 技術分享】如何透過 AWS WAF 為 Amazon Bedrock AgentCore 建立安全防線？

Advanced (300)Amazon Bedrock AgentCoreAmazon VPCAWS WAF

Selective Timestep Weighting and Advantage-Based Replay for Sample-Efficient Diffusion RLHF

📌 【研究】Diffusion RLHF 效率突破：透過時間步權重與重播機制提升 6 倍樣本效率

cs.LGcs.AIcs.CV

Show HN: Microsoft releases Flint, a visualization language for AI agents

📌 Flint：AI 代理高階視覺化語言

TencentCloud/CubeSandbox

📌 TencentCloud 極速沙箱

Rust

剛剛，全球首個具身專屬的MoE影片模型，開源了！

📌 【螞蟻靈波】首個具身專屬 MoE 影片模型 LingBot-Video 開源：讓機器人學習真實物理規律

资讯首页轮播LingBot-Video具身智能

WildCity: A Real-World City-Scale Testbed for Rendering, Simulation, and Spatial Intelligence

📌 WildCity：城市尺度多模態資料集，推動空間智慧研究

螞蟻靈波開源LingBot-Video，全球首個面向具身的影片基模來了！

📌 螞蟻靈波開源LingBot-Video，全球首個具身影片基模

资讯蚂蚁灵波

Separating signal from noise in coding evaluations

📌 【OpenAI 最新分析】SWE-bench Pro 基準測試存在可靠性問題

Research

Attending to Multimodal Generation One Token at a Time

📌 多模態 LLM 的注意力動態：視覺與文本的此消彼長

3D HAMSTER: Bridging Planning and Control in Hierarchical Vision Language Action Models through 3D Trajectory Guidance

📌 3D HAMSTER：用深度視覺語言產生實際 3D 軌跡，串接規劃與控制

CanvasAgent: Enabling Complex Image Creation and Editing via Visual Tool Orchestration

📌 CanvasAgent：用視覺工具串聯，搞定複雜繪圖工作流

Bibby AI: An Editor-Native Agentic Platform for Academic Research, Writing, and Publishing

📌 Bibby AI：整合型編輯原生學術寫作代理平臺

Citrix NetScaler adds MCP Gateway to secure and govern AI agent traffic

📌 NetScaler 新增 MCP Gateway，守護 AI Agent 連線安全

Cross-Space Distillation: Teaching One-Step Students with Modern Diffusion Teachers

📌 HuggingFace 論文】用 VAE 空間對齊，實現一步生成擴散模型

Flex-Forcing: Towards a Unified Autoregressive and Bidirectional Video Diffusion Model

📌 Flex-Forcing：同時支援自回歸與雙向生成的影片 Diffusion 模型

GitLost: We Tricked GitHub's AI Agent into Leaking Private Repos

📌 Noma Labs 揭露 GitHub Agentic Workflows 重大漏洞

HunyuanOCR-1.5: Making Lightweight OCR VLMs Faster and Better

📌 騰訊 HunyuanOCR-1.5：輕量視覺語言模型的效率進化

Gemma 4 Technical Report

📌 Gemma 4：多模態與推理能力的新平衡

Layer-wise Cross-Lingual Depression Detection from Speech: Analysis with Contrastive Alignment

📌 跨語言抑鬱偵測：對比對齊讓英語與華語共享臨床空間

MUFG aims to become AI-native with OpenAI

📌 MUFG 搭配 ChatGPT Enterprise 打造 AI 原生企業

RynnWorld-4D: 4D Embodied World Models for Robotic Manipulation

📌 RynnWorld‑4D：從單張 RGB‑D 圖片產生 4D 互動世界模型，提升機械手臂操作效能

Light-Omni: Reflex over Reasoning in Agentic Video Understanding with Long-Term Memory

📌 Light-Omni：雙重記憶狀態，讓 AI 看懂影片不再反覆推理

RynnWorld-Teleop: An Action-Conditioned World Model for Digital Teleoperation

📌 用生成式世界模型打造數位遙控，加速機器人零樣本遷移

SkillOpt-Lite: Better and Faster Agent Self-evolution via One Line of Vibe

📌 SkillOpt-Lite：一行指令加速 Agent 自我進化

TREK: Distill to Explore, Reinforce to Refine

📌 TREK：用蒸餾做探索，強化來精煉

TurnOPD: Making On-Policy Distillation Turn-Aware for Efficient Long-Horizon Agent Training

📌 TurnOPD：讓 On-Policy 蒸餾具備「回合感知」能力

Vision as Unified Multimodal Generation

📌 Vision as Unified Multimodal Generation

We charge $10k a week to delete AI-generated code

📌 週費一萬美金刪除 AI 程式碼？清理 Vibecoded 專案的現實策略

A Single Neuron Is Sufficient to Bypass Safety Alignment in Large Language Models

📌 單一神經元即可繞過大型語言模型的安全對齊

anthropics/skills

📌 【Anthropic 開源】Claude 的「Skills」機制：讓 AI 透過動態載入資源完成專業任務

Python

Build an AI-powered AWS support companion with Amazon Bedrock AgentCore

在處理 AWS 基礎設施事件時，工程師通常需要：

Advanced (300)Amazon BedrockAmazon Bedrock AgentCoreAmazon CloudWatch

dOPSD: On-Policy Self-Distillation for Diffusion Language Models

📌 Diffusion LLM 的新突破：透過 On-Policy 自蒸餾強化推理能力

Develop Humanoid Robot Policies End-to-End with NVIDIA Isaac GR00T

📌 NVIDIA Isaac GR00T：端到端人形機器人政策開發

RoboticsSimulation / Modeling / DesignAI Foundation ModelsHumanoid Robots

Expanding Managed Agents in Gemini API: background tasks, remote MCP and more

📌 Gemini API 新增 Managed Agents 功能：支援背景執行、遠端 MCP 與自訂函式

Developer tools

kyutai-labs/pocket-tts

📌 【kyutai-labs】Pocket TTS：無需 GPU 也能實現低延遲、即時語音合成

Python

Bridging Interleaved Multi-Modal Reasoning as a Unified Decision Process

Bridging Interleaved Multi-Modal Reasoning as a Unified Decision Process

LensVLM: Selective Context Expansion for Compressed Visual Representation of Text

📌 LensVLM：以選擇性展開提升壓縮文字影像的視覺語言模型效能

MT-EditFlow: Reinforcement Learning for Multi-Turn Image Editing with Flow Matching

📌 MT-EditFlow：用 RL 與 Flow Matching 提升多輪影像編輯

NVIDIA Vera CPU Boosts AI Factory Throughput to Accelerate Agentic Workloads

NVIDIA 推出的 Vera CPU 採用單片計算晶片，擁有 88 個 Olympus 核心、統一快取與 Scalable Coherency Fabric。在全槽負載下，單核心持續效能較傳統 x86 資料中心 CPU 提升 1.8 倍；峰值載入延遲降低 40%；每核記憶體頻寬提升超過 3 倍，且…

Agentic AI / Generative AITop StoriesAI AgentAI Factory

OpenAI Releases GPT-Realtime-2.1 and GPT-Realtime-2.1-mini for Low-Latency Voice Agents in the API

📌 OpenAI 推出 GPT-Realtime-2.1 系列：低延遲語音代理現在能「思考」後再說話

Agentic AIAudio Language ModelLanguage ModelLarge Language Model

Safety Testing LLM Agents at Scale: From Risk Discovery to Evidence-Grounded Verification

📌 如何大規模測試 LLM Agent 的安全性？Vera 框架將風險發現與驗證系統化

Teaching models to forget: Selective unlearning with Amazon Nova

組織在使用基礎模型時，常遇到模型內建的內容審核機制過度拒絕（over‑deflection），導致合法且具業務價值的內容被擋下。AWS 推出 Amazon Nova Customizable Content Moderation Settings (CCMS)，其核心是 Reverse Direct…

Amazon SageMakerAnnouncementsFoundational (100)

Tencent Releases Hy3: An Open 295B Mixture-of-Experts (MoE) Model with 21B Active Parameters and 256K Context

📌 【Tencent 最新釋出】Hy3 開源 295B MoE 模型：256K 長上下文與 MTP 加速推理

Agentic AILanguage ModelLarge Language ModelMachine Learning

SeKV: Resolution-Adaptive KV Cache with Hierarchical Semantic Memory for Long-Context LLM Inference

📌 SeKV：透過層級化語義記憶體，突破長文本推論的 KV Cache 瓶頸

TencentCloud/TencentDB-Agent-Memory

📌 TencentDB Agent Memory：符號化短期記憶＋分層長期記憶，降低 Token 用量並提升任務成功率

TypeScript

The ‘first’ AI-run ransomware attack still needed a human

📌 首起 AI 勒索軟體攻擊仍需人類介入

Securityransomwaresysdig

從零開始玩轉迴圈 (Getting started with loops)

📌 Claude Code 迴圈分類與使用指南

Weblica: Scalable and Reproducible Training Environments for Visual Web Agents

📌 Weblica：可擴充套件且可重現的視覺網頁代理訓練框架

A global workspace in language models

📌 【Anthropic 最新研究】Claude 內部存在「全域工作空間」：模型在不輸出文字前就在「思考」

Research

Embodied.cpp: A Portable Inference Runtime of Embodied AI Models on Heterogeneous Robots

📌 Embodied.cpp：在異構機器人上以 C++ 部署具身 AI 模型的可移植推論執行環境

Deploying Multi-Turn RL Infrastructure for Amazon Nova on Amazon SageMaker HyperPod

📌 【AWS 分享】如何利用 SageMaker HyperPod 部署 Amazon Nova 的多回合 RL 基礎建設

Amazon SageMaker AIExpert (400)Technical How-to

Getting Started with Hugging Face ML Intern: Your First ML Agent

📌 【Hugging Face 新工具】ML Intern：用自然語言描述，讓 AI 幫你完成模型訓練全流程

jamwithai/production-agentic-rag-course

📌 【GitHub 趨勢】別再只用向量搜尋：從基礎關鍵字到生產級 Agentic RAG 的實作路徑

Python

LeRobot v0.6.0: Imagine, Evaluate, Improve

📌 【HuggingFace】LeRobot v0.6.0：引入世界模型與自動化回饋迴路，縮短機器人學習路徑

LlamaIndex ‘legal-kb’: Agentic Retrieval over Index v2 with retrieve, find, read, and grep Tools

📌 LlamaIndex 推出 legal-kb：用「檔案系統操作」實現 Agentic 檢索

Agentic AIOpen Source

Meituan Releases LongCat-2.0: A 1.6T-Parameter Open MoE Model with Native 1M Context and LongCat Sparse Attention

📌 美團發布 LongCat-2.0：1.6T 引數 MoE 模型，支援 1M 上下文

Agentic AIAI InfrastructureSoftware Engineering

microsoft/qlib

📌 【Microsoft】RD-Agent：用 LLM 多代理自動化量化投資的因子挖掘與模型最佳化

Python

OrbitQuant: Data-Agnostic Quantization for Image and Video Diffusion Transformers

📌 OrbitQuant：讓 Diffusion Transformer 量化不再需要針對每個時間步反覆校正

Path-Constrained Mixture-of-Experts

📌 【Apple 研究】PathMoE：將 MoE 從「單層路由」轉向「路徑約束」提升效能

Revisiting ASR Error Correction with Specialized Models

📌 【Apple 研究】捨棄 LLM？用輕量化 seq2seq 模型更精準地校正 ASR 錯誤

Qwen’s Former Lead on What Hybrid Thinking Got Wrong — and Why He Now Backs Agents

📌 【前 Qwen 技術領袖觀點】從訓練模型轉向訓練 Agent：Hybrid Thinking 的挑戰與未來

Agentic AIAI AgentsJunyang Lin

Securing the AI Agent: A Unified Framework for Multi-Layer Agent Red Teaming

📌 AI Agent 安全防禦：AI-Infra-Guard 提出多層次紅隊測試框架

Structured PDF-to-JSON: A Guide to Open-Source Extraction Models in 2026

📌 Structured PDF-to-JSON：2026 年開源抽取模型指南

Language ModelOCROpen SourceSoftware Engineering

Synthetic Sciences Releases OpenScience: An Open-Source, Model-Agnostic AI Workbench for Machine Learning, Biology, Physics, and Chemistry Research

📌 OpenScience 開源科研 AI 工作臺

Agentic AIAI InfrastructureLanguage ModelLarge Language Model

TauricResearch/TradingAgents

📌 【開源專案】TradingAgents：建構多代理 LLM 金融交易框架

Python

Scaling Properties of Continuous Diffusion Spoken Language Models

📌 【Apple ML 研究】連續擴散模型能否打破語音語言模型的尺度瓶頸？

Training Gemma-3 for Structured Mathematical Reasoning with Tunix GRPO, LoRA Adapters, and GSM8K Rewards

📌 【實作教學】結合 GRPO 與 LoRA 強化 Gemma-3 的結構化數學推理能力

Language ModelLarge Language ModelMachine LearningTutorials