LLM 研究知识库

概念 MoshiRAG: Asynchronous Knowledge Retrieval for Full-Duplex Speech Language Models 首个将 RAG 集成到全双工语音语言模型的系统。核心思想：利用全双工对话中"回复开头"到"关键信息出现"之间的天然时间差（keyword delay），在此期间异步触发外部知识检索，检索结果在关键信息到达前注入模型，从而在不破坏实时交互的前提下大幅提升事实性。基于 moshi 7B 模型，检索后端可插拔（Gemma 3 27B / GPT-4.1 / Tav... TL;DR问题定义系统架构 2026-06-07 实体 Qwen3.5 Qwen3.5 是阿里巴巴通义千问团队发布的下一代基础模型系列。首个开源权重模型为 Qwen3.5-397B-A17B。 概述架构创新推理优化支持 2026-05-12 概念 EAGLE (Speculative Decoding) EAGLE 是一种 外挂模型 (External Draft Model) 的投机解码技术。它训练一个轻量级的 Draft Model 来“模仿”Target Model 的生成风格，从而并行生成多个候选 token，再由 Target Model 验证。 概述核心机制版本演进 2026-05-12 概念 Multi-Token Prediction (MTP) MTP (Multi-Token Prediction) 是一种 原生内置 (Native) 的投机解码 (Speculative Decoding) 技术。与外挂的 Draft Model 不同，MTP 将预测头直接嵌入在 Transformer 的中间层中，在预训练阶段就学习预测多个未来的 token。 概述核心机制优势与劣势 2026-05-12 概念 Agent Safety Agent safety covers methods for preventing tool-using and autonomous agents from executing harmful, unsafe, or policy-violating behavior across planning, memory, tool use, and mult... Related 2026-05-10 概念 Agentic Research Ideas Agentic research ideas are research directions where agents help formulate hypotheses, design experiments, operate tools, and iterate on findings. Related 2026-05-10 概念 Agentic Workflows Agentic workflows are manually or automatically designed chains of agent steps, tools, roles, and verification loops. Related 2026-05-10 概念 AgentPoison AgentPoison refers to backdoor or poisoning attacks against agent memory, where malicious content is inserted into persistent state and later influences agent decisions. Related 2026-05-10 概念 ASR Benchmarks ASR benchmarks evaluate automatic speech recognition quality across languages, accents, noise conditions, and domain-specific speech. Related 2026-05-10 概念 Complementary Learning Systems Complementary Learning Systems is a theory of memory that separates fast episodic acquisition from slower consolidation into structured long-term knowledge. Related 2026-05-10