搜索结果：AI大模型

This research paper reveals new indirect prompt injection attack vectors that enable adversaries to remotely exploit LLM-integrated applications by injecting malicious prompts into retrieved data, demonstrating practical attacks against systems like Bing Chat and highlighting significant security vulnerabilities in current LLM deployments. 原文翻译： 本研究论文揭示了新的间接提示注入攻击向量，攻击者可通过在检索数据中注入恶意提示来远程利用LLM集成应用，展示了针对Bing Chat等系统的实际攻击，并突显了当前LLM部署中的重大安全漏洞。

This guide explores how Large Language Models (LLMs) can accelerate GPU kernel optimization, bridging the gap between research algorithms and production-ready implementations. It covers the computational foundations of matrix multiplication, the mathematical framework from Strassen to tensor decomposition, and DeepMind's evolution from AlphaTensor to AlphaEvolve, providing a technical roadmap for self-improving AI infrastructure. 原文翻译： 本指南探讨了大型语言模型（LLM）如何加速GPU内核优化，弥合研究算法与生产就绪实现之间的差距。它涵盖了矩阵乘法的计算基础、从Strassen到张量分解的数学框架，以及DeepMind从AlphaTensor到AlphaEvolve的演进历程，为自改进AI基础设施提供了技术路线图。

Dive into LLMs is a comprehensive, practice-oriented tutorial series for large language models, developed from Shanghai Jiao Tong University's AI Security Technology course with Huawei Ascend support. It covers the full technical stack from fine-tuning to security alignment, offering complete learning resources and domestic hardware adaptation. 原文翻译： Dive into LLMs 是由上海交通大学《人工智能安全技术》课程拓展而来的大模型编程实践教程系列，获得华为昇腾社区支持。它以实践为导向，覆盖从微调到安全对齐的全技术栈，提供完整学习资源并适配国产硬件。

RAG (Retrieval-Augmented Generation) is a mainstream LLM application architecture that enhances answer accuracy by retrieving relevant information from external knowledge bases before generation, addressing issues like knowledge cutoff and hallucinations. 原文翻译： RAG（检索增强生成）是目前主流的LLM落地架构，通过在生成答案前从外部知识库检索相关信息，显著提升回答准确性，有效解决知识截止日期和幻觉问题。

This article explores how large language models (LLMs) function as document completion engines and demonstrates how to build practical applications by mapping between user and document domains, using GitHub Copilot's prompt engineering pipeline as a detailed case study. 原文翻译： 本文探讨了大型语言模型（LLM）如何作为文档完成引擎工作，并通过在用户领域和文档领域之间建立映射来构建实际应用，以GitHub Copilot的提示工程流程作为详细案例研究。

This article analyzes the winning solutions from the Enterprise RAG Challenge, detailing the architectures, models, and experimental approaches used by top performers like Ilya Rice and Emil Shagiev. It provides insights into effective RAG patterns, including dense retrieval, LLM reranking, router patterns, and structured outputs. 原文翻译： 本文分析了企业RAG挑战赛的获胜解决方案，详细介绍了Ilya Rice和Emil Shagiev等顶级选手使用的架构、模型和实验方法。文章深入探讨了有效的RAG模式，包括密集检索、LLM重排序、路由模式和结构化输出。

Amazon's DGL-KE tools optimize knowledge graph embedding for parallel computing, achieving 2-5x efficiency gains through graph partitioning, negative sampling, relation partitioning, and overlapping computations. 原文翻译： 亚马逊的DGL-KE工具通过图分区、负采样、关系分区和计算重叠等优化技术，显著提升了知识图谱嵌入在并行计算环境中的效率，实现了2-5倍的性能提升。

Librarium is a command-line tool that parallelizes research queries across multiple search and deep-research APIs, collecting and normalizing results into structured output. It features 18 built-in providers across three tiers (deep-research, AI-grounded, raw-search) and supports custom providers. 原文翻译： Librarium 是一个命令行工具，可将研究查询并行分发到多个搜索和深度研究 API，收集并规范化结果，输出结构化数据。它包含 18 个内置提供商，分为三个层级（深度研究、AI 基础搜索、原始搜索），并支持自定义提供商。

Structured generation in LLMs consistently improves performance on the GSM8K benchmark, with up to 70%+ gains, and offers additional benefits like prompt consistency and thought-control. 原文翻译： LLM中的结构化生成在GSM8K基准测试中持续提升性能，最高可达70%以上的增益，并带来提示一致性和思维控制等额外优势。

This comprehensive 2026 ranking analyzes major LLMs across reasoning, coding, math, agentic, software engineering, and chat benchmarks, providing detailed performance scores and pricing comparisons for technical professionals. 原文翻译： 这份全面的2026年排名分析了主要大语言模型在推理、编码、数学、代理、软件工程和聊天基准测试中的表现，为技术专业人士提供详细的性能得分和定价对比。