推理吞吐的分水岭在批处理调度：静态 batching 为什么浪费（长短请求互相等），continuous batching 的 iteration 级调度怎么做到完成的即走、新请求随到随插，它与 PagedAttention 的共生关系，以及批大小与等待窗口的取舍。

Hierarchical language generation separates continuous semantic modeling from discrete text realization, but its effectiv…

A main promise of looped language models is depth-adaptive inference. By looping a block of shared layers a variable num…

Real-time long-video generation requires continuous output with consistent appearance and natural motion. Existing chunk…

Learning the continuous-time evolution of nonlinear dynamical systems relies on trajectory-based objectives, which requi…

Multimodal large language models (MLLMs) are rapidly evolving toward continuous audio--visual reasoning, creating an urg…

Continuous asynchronous replanning is essential for real-time generative robot policies, but independent stochastic init…

We present Multimodal Flow, a fully continuous generative model of language and vision. Most unified multimodal models e…

System performance, efficient infrastructure scaling and continuous software optimization are key levers that determine …

Natural interaction in digital and physical environments requires continuous perception and timely responses. Spoken dia…

Recent advances in wearable sensing enable continuous monitoring of physiological and behavioral signals, yet existing b…

We present a continuous, population-scale measurement record of autonomous language-model trading agents operating in pr…

从引擎到服务的最后一公里——拆解前缀感知路由、cache-aware 负载均衡、LoRA 热加载与模型热切换的生产实现。

同一个输入为什么每次输出不同：采样随机性、浮点非结合性、批处理与推理栈三层来源，以及 seed 的 best-effort 本质、温度 0 的陷阱与可复现工程实践。

At a glance Formal methods for everyday software.** PDD Kit helps you turn your intent into a formal spec, then proves t…

深度伪造篇拆解了攻击的技术，这篇拆解防御的产业：活体检测的攻防军备、身份验证的多因子重构、以及「人脸不再等于身份」之后，金融与政务的 KYC 体系怎么重建。

一次智能体响应的延迟由预填充、解码、工具执行与多轮串联构成，本文拆解延迟结构，梳理 prompt 缓存、并行化、语义缓存与推理侧加速四条优化路径，并给出延迟预算分配法。

LLM 推理服务的伸缩比普通服务难：副本启动要加载几十 GB 权重（冷启动分钟级）、显存装不下就想不加、流量是尖峰型。本篇拆解推理服务伸缩的三大难题与对策——模型加载的冷启动、基于指标的 KPA 伸缩、以及「预热」与「错峰」的运营手段。

一张 A100 想给十个团队用、一张卡想同时跑三个小模型——GPU 共享的三条技术路线：MIG 的硬件级切分、vGPU 的虚拟化直通、时分复用的软件调度。本篇拆解三种路线的机制、隔离性与适用场景，以及「共享 GPU」的噪声与公平问题。

同一张 GPU，用 naive 推理脚本与用 vLLM，吞吐可以差 10 倍以上——差距的来源是两项工程创新：PagedAttention（显存的分页管理）与连续批处理（请求的动态编排）。本篇拆解这两项技术的机制、vLLM 生态的位置、以及「什么时候该用推理引擎、什么时候杀鸡不必用牛刀」。

大模型价格战的背后（本库《价格战》），站着一条被低估的产业链：推理优化产业——量化、编译器、KV 缓存优化、投机解码、专用推理引擎。它不训练模型，却决定了模型的交付成本，是「价格战能打多低」的真正决定者。本文画一张推理优化产业的地图，讲清每一格的手艺与玩家。

思维链让模型把思考写成文字，但文字真的是推理的必要载体吗？COCONUT（链式连续思维）让模型直接在隐藏状态里「想」，跳过语言化的中间步骤——效率与能力的权衡揭开了「语言是不是思维的枷锁」这条研究线。本篇拆解隐式推理的动机、COCONUT 的机制、潜空间推理的进展与局限。

Despite recent advances in CUAs, their development remains constrained by the limited diversity, controllability, and re…

AI-generated videos often depict people within detailed scenes while leaving their small facial regions poorly resolved.…

Personal AI agents are increasingly capable of handling long-horizon tasks across applications, yet users often need to …

Most robot policies, whether trained by imitation learning (IL) or reinforcement learning (RL), map observations to acti…

As publications and products on time series forecasting (TSF) proliferate, researchers face a growing burden of analyzin…

Predicting future physical states stands as a central challenge in embodied intelligence. Existing video world models le…

On-Policy Self-Distillation (OPSD), which aligns a student model's token-level distribution with that of its self-teache…

Oct 24, 2026 Nano Banana Pro prompted by THE DECODER Key Points After a security incident in September, OpenAI now train…

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

To help developers rapidly build world models, which help physical AI systems understand and simulate the real world, NV…

General availability of the instances is coming next month, with planned expansion of availability to Europe — bringing …

Vision-language-action (VLA) models have shown strong promise for generalist robotic manipulation, but their inference c…

The Nordic countries have become the European leader in unicorn startup density, producing the most emerging companies v…

Companies in France are deploying agentic computing across industries spanning space, financial services, the public sec…

★36,402、Apache-2.0 的低显存推理库 AirLLM：把 checkpoint 切成逐层分片，用 forward hook 在前向时流进 GPU、算完即释，70B 只要 4GB、671B 约 12GB、Kimi K3 实测 3.72GB。拆解 meta device + 2GiB pin 上限 + MoE 按专家流式的取舍，以及社区实测 5–35 秒/token 的真实代价。

论文提出分块流式几何基础模型 Drive3R，通过分块因果注意力解决自动驾驶长序列多视图3D重建中GPU内存与全局注意力二次复杂度的瓶颈。

Rather than downloading another app, a growing number of agents can simply be texted like an ordinary person. You text i…

Multimodal language models (MLLMs) exhibit promising spatial reasoning capabilities, yet the internal mechanisms support…

Diffusion and autoregression (AR) have long been seen as different categories of generative models, with diffusion speci…

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

The story is a collaboration between MIT Technology Review and Aventine, a non-profit research foundation that creates a…

In this work, we show that a single Transformer block, applied recurrently, can match the accuracy of a full-depth visio…

研究团队推出RobotWorld模拟测试平台，含84项覆盖操作、移动操作、运动控制、驾驶与飞行控制的任务，通过交互预算与可执行成功检验评估智能体将指令转化为物理任务执行的能力，并分析哪些能力可迁移、哪些缺口导致无法可靠完成。

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

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

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

多请求并发时 KV Cache 的预留式分配让 GPU 显存大量空转、吞吐卡死。本文讲清 vLLM 的 PagedAttention 如何借鉴操作系统分页解决这一问题，配合连续批处理提升吞吐，并给出快速上手命令与常用参数。

论文提出TIDES，一种选择性SSM变体，通过将输入依赖从时间步长转移到对角状态矩阵，使模型在保持选择性的同时原生处理不规则时间戳。