This article introduces practical methods for evaluating AI agents operating in real-world environments. It explains how to combine benchmarks, automated evaluation pipelines, and human review to ...

The Farmer Was Replaced is part programming lesson and part automation title, and it has players program a drone to automate tasks on a farm.

人工智能领域正经历一场从“模型中心”向“Agent系统”的范式转移。传统AI系统以单一模型为核心，通过输入数据、模型计算、输出结果的线性流程完成任务，其资源消耗集中于GPU算力。但随着AI Agent技术的突破，计算模式发生根本性变化——系统开始整合GPU、CPU、API接口、存储设备和网络资源，形成多维度资源协同的复杂架构。 以典型任务为例，现代AI ...

多模态大模型在代码能力上进步惊人，但在基础视觉任务上却频繁失误。 活跃在AGI基础研究前沿的技术团队UniPat AI构建了一个极简的视觉智能体框架——SWE-Vision，让模型可以编写并执行Python代码来处理和验证自己的视觉判断。在五个主流视觉基准测试中，SWE-Vision均达到了当前最优水平。 模型看得见，却没法精确处理 多模态大模型的代码能力在过去一年取得了惊人进展——独立搭建项目、 ...

3月16日消息，小米AI实验室研究员罗福莉，也就是很多人口中的“天才少女”，又发论文了。论文名叫ARL-Tangram: Unleash the Resource Efficiency in Agentic Reinforcement ...

To address these shortcomings, we introduce SymPcNSGA-Testing (Symbolic execution, Path clustering and NSGA-II Testing), a ...

论文名叫ARL-Tangram: Unleash the Resource Efficiency in Agentic Reinforcement Learning。作者之一，就是罗福莉。 要用GPU去做模型推理，要用CPU去执行代码，要用API去处理搜索、数据库，可能还要用浏览器进行网页操作 ...

多模态大模型在代码能力上进步惊人，但在基础视觉任务上却频繁失误。UniPat AI 构建了一个极简的视觉智能体框架 ——SWE-Vision，让模型可以编写并执行 Python 代码来处理和验证自己的视觉判断。在五个主流视觉基准测试中，SWE-Vision 均达到了当前最优水平。 多模态大模型的代码能力在过去一年取得了惊人进展 —— 独立搭建项目、排查 bug、完成复杂重构，表现已可比肩资深工程师 ...

A Hong Kong court has ruled that two Tiananmen vigil activists have a case to answer over calls to “end one-party rule” in China in a subversion trial under the Beijing-imposed national security law.

Infosecurity spoke to several experts to explore what CISOs should do to contain the viral AI agent tool’s security vulnerabilities ...

Testing is where Thailand's AI adoption often pays off quickly, because it reduces waiting. AI can draft unit tests from code, suggest regression ...

Researchers show AI can learn a rare programming language by correcting its own errors, improving its coding success from 39% to 96%.