模型与评测

From Tokens to Semantics: Leveraging Complementary Signals for Hallucination Detection in Black-Box LLMs

When LLMs support public-facing or high-stakes workflows, missed fabrications can harm users and institutions, while false alarms consume limited human-review capacity. When no tru

Gemini Omni 1.1 Flash lets you build with more control

GeoJSON Map Viewer

Tool: GeoJSON Map Viewer I was helping Natalie gather some maps of local political boundaries (for the Granada Community Services District and the Midcoast Community Council ) and

GitHub Skills

Glance：单次细粒度图像描述

一次前向做细粒度 caption，配 SimLoss。

Go grandmaster Shin defeats AI KataGo in historic human victory

Google AI 团队分享如何为 LLM-as-a-Judge 评测编写可靠的评分标准

Google AI 团队发布教程，讲解如何为 LLM-as-a-Judge 评测编写可靠的布尔式评分标准，指出模糊提示会导致评估不一致和浪费 token。文中给出四条经验：问题保持原子化且互不重叠、只让评判模型评估客观事实（可用 RFC 2119 术语如 MUST 表述）、只评 prompt 中明确要求的内容、用专家标注的 golden set 校准评判模型

Google DeepMind 发布 Gemini 3.8 Flash 与 3.8 Flash Cyber 两款新模型

Google DeepMind 发布 Gemini 3.8 Flash 与 3.8 Flash Cyber。 🔗 阅读原文 via AIHOT · https://aihot.virxact.com/items/cmtkbdbti01n8roz5k5kt1g98

Google Gemini's new agent-based video analysis cuts token usage by up to 88 percent

Google is adding agent-based video analysis to Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite. Instead of scanning videos frame by frame at a fixed rate, the model decides on its

Google avoids a breakup of its ad tech business

Google 让 MrBeast 带着 AI 进荒野

Gemini + Fitbit 的内容合作。

Gradium 新默认 TTS：难例 81%、首音频 216ms

Gradium AI 发布新默认 TTS，宣称难例通过率 81.0%、首音频延迟约 216ms，适合实时配音。

H3-World：语言理解变成世界控制

把语言理解接到世界控制。

Harness-of-Harness：多日自主开发

连续多日软件开发，套件还能自我改进。

HarnessDev：模型自造 Agent 套件

问 LLM 能否创建并进化自己的 agent harness。

HiPoly: a hierarchical polymer-native AI framework for property prediction and generative design

Polymeric materials are central to modern technologies, with applications ranging from energy to health and transportation. Although AI has made significant advances in materials d

Hugging Face 发布 @huggingface/kernels，提供 207 个 WebGPU 内核用于浏览器本地 AI 推理

Hugging Face WebAI 团队发布 @huggingface/kernels 库及 207 个以独立仓库形式托管在 Hub 上的 WebGPU 内核（Apache-2.0），每个内核带 manifest、正确性测试、基准用例和 WGSL 着色器模板。 🔗 阅读原文 via AIHOT · https://aihot.virxact.com/ite

I tried to predict which jobs AI replaces. The useful part is where it's wrong

Import AI 471: Why Hugging Face worries me; space mining; FIve Eyes on AI

Welcome to Import AI, a newsletter about AI research. Import AI runs on arXiv, cappuccinos, and feedback from readers. If you’d like to support this, please subscribe. Subscribe no

Introducing Hy4 Preview

Introducing Hy4 Preview New open weight text input (no vision) LLM from Chinese company Tencent today: 770B total parameters, 49B active parameters, 1M token context window, 1.56TB

Introducing wrapture

Introducing wrapture New from Graham Dumpleton (of wrapt , mod_wsgi, and New Relic's Python agent fame), who describes Wrapture as taking the monkeypatching ideas from wrapt and ex

Issues Plan and track work

Just a rumour of a bug is enough to find a security exploit these days

Just a rumour of a bug is enough to find a security exploit these days Anil Madhavapeddy is a professor of computer science at Cambridge and a core maintainer of the OCaml compiler

Kirin：野外视频生成动物运动

从 in-the-wild 视频生成动物运动。