同一个模型，temperature 从 0.2 调到 1.2，可以从「背答案的图书馆员」变成「喝了两杯的诗人」——模型一个参数都没变，变的只是从概率分布里挑词的规则。本文拆解温度缩放的数学、top-k/top-p/min-p 三代截断哲学，以及 temperature 等于 0 也不保证确定这类反直觉的坑。

模型每步输出的不是「下一个字」而是一张概率表，Temperature 与 Top-p 决定怎么从表里抽签。本文讲清温度如何改变分布锐度、Top-p 如何动态截断候选、重复惩罚如何抑制复读，并给出代码、问答、写作、Agent 四类场景的参数对照表。

Solutions based on large language models (LLMs) often rely on temperature sampling to improve accuracy and stability by …

!(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…

硬件初创公司Natura发布99美元智能戒指Interface，用户按压手指即可调用AI代理完成任务，同时兼具心率、HRV、睡眠和活动等健康监测功能。

JetBrains 推出面向编码智能体的开源模型 Mellum2.1，总参数 12B、每 token 激活 2.5B，主要通过在真实软件环境中的强化学习带来提升，以 Apache 2.0 协议在 Hugging Face 发布，可自托管运行。

2024 年 9 月 o1 用「先想一会儿」在 AIME 上打出 83% 对 13%，2025 年 1 月 DeepSeek-R1 把整套方法开源复现，到 2026 年 10 月，推理档位已成各大模型 API 的标准旋钮。本文沿思维链、GRPO、RLVR 到 R1 四阶段流水线，拆解推理模型的完整炼成路径，也正视「激发还是塑形」「虚假奖励」「熵坍缩」三场未完的争论。

!(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…

OpenAI、Anthropic、Google 三家官方文档一致把推理模型指向数学、编程、复杂调试与长程智能体任务，而简单分类、低延迟与高吞吐场景不值得开思考。本文梳理三家的思考参数与档位选择，算清「思考 token 按输出计费」的成本账，给出一套 effort 档位取舍与调参的实用框架。

把 LLM 输出接进程序，坏 JSON 是头号工程痛点。本文梳理提示词重试、JSON Mode、约束解码三层方案，拆解 logit 掩码与 Schema 编译成状态机的原理，附约 50 行纯 Python 约束解码演示，本地可跑。

小模型学的是大模型的「能力」而非权重：讲清 Hinton 软标签与温度 τ 的暗知识原理、白盒与黑盒两条蒸馏路线、DeepSeek-R1 用 80 万样本蒸出 Qwen/Llama 小模型的成绩，以及学生上限与闭源 ToS 等边界。

输出流畅、语气自信，内容却是编的——幻觉是当前大模型架构的固有属性，不是能修掉的 bug。本文拆解幻觉生成的三层机理、最容易被放大的三类场景，给出检测思路与从 RAG 到提示词设计的缓解清单。

想在笔记本上跑大模型，Ollama 是门槛最低的路径之一。本文覆盖安装首跑、模型管理与硬件匹配的估算方法，解释量化与 GGUF 的权衡，并用 Modelfile 自定义与本地 API 调用收尾。

Agent 工程系统学习第 1 章：从 HTTP 协议层讲透 LLM API——Chat Completions 协议与 role 语义、Function Calling 的"模型选择/代码执行"分工与三大常见错误、SSE 流式手写解析器（含 tool_calls 分块拼接）、token 计量与前缀缓存工程、重试/超时/幂等的错误分类纪律、多模态输入成本。附零框架多轮工具 Agent 实现作业。

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

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

NVIDIA has released Kumo Tabular, a new family of tabular foundation models (TFMs) for classification and regression. If…

Large language models (LLMs) excel at structured, verifiable tasks, but their low-entropy bias can produce homogeneous a…

Liquid AI has announced LFM2.5-VL-3B-DSpark, an experimental speculative-decoding draft model for its LFM2.5-VL-3B visio…

The US government wants to spend $30.3 million over the next five years on an improved form of lie detector, according t…

BottleCap AI has released ThinkingCap-Qwen3.8-27B, the second model in its ThinkingCap series. It is a fine-tune of the …

This is today's edition of* *The Download*,*our weekday newsletter that provides a daily dose of what's going on in the …

Nokia’s applied research team has open-sourced AnyJev, a Python library that turns an open LLM into a decision model. It…

Kyutai has released **Voice of Reason**, 2 open-weight speech-to-speech models that solve math problems out loud. Both s…

Google 开源的智能体执行编排运行时 AX（当日涨星 2,324、★7,482、Apache-2.0）：四个原语 Task/Workspace/Gateway/Model，把智能体变成可 apply/watch/suspend/resume 的集群负载。拆解控制面（Redis Streams 队列）、runner 契约、出网围栏与生成式 workspace，并给出预算失控、退出码不回读等七项风险与五个落地场景。

Clean energy isn’t hard to come by, but the pace of large-scale adoption has historically been slow due to bottlenecks —…

This week two conversations about AI safety went viral that demonstrate just how hard it is to discern AI fact from fict…

Sleeping on the Orion is like flipping your pillow to find “the cool side,” except that your entire bed is the cool side…

Google Research has introduced Retrieve-for-Train (R4T), a framework for search that returns coherent, diverse result se…

The AI boom is becoming a materials challenge. As AI pushes computing into new territory, the materials behind that infr…

Vibe design agents turn natural-language briefs into rendered interfaces and frontend code. Yet a useful design agent sh…

Reinforcement learning (RL) has become central to post-training of large language models. Recent advances in RL for Mixt…

Tokenizer 的工作机制：BPE 分词、上下文窗口与计费逻辑