<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>扩散语言模型 on AI内参</title>
    <link>https://www.neican.ai/tags/%E6%89%A9%E6%95%A3%E8%AF%AD%E8%A8%80%E6%A8%A1%E5%9E%8B/</link>
    <description>Recent content in 扩散语言模型 on AI内参</description>
    <generator>Hugo</generator>
    <language>zh-cn</language>
    <lastBuildDate>Fri, 12 Dec 2025 19:24:26 +0800</lastBuildDate>
    <atom:link href="https://www.neican.ai/tags/%E6%89%A9%E6%95%A3%E8%AF%AD%E8%A8%80%E6%A8%A1%E5%9E%8B/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>AI Agent：一场从「Token内卷」到「星际计算」的狂飙，人类，准备好了吗？</title>
      <link>https://www.neican.ai/newspaper/2025-12-12-ai-agent-token-/</link>
      <pubDate>Fri, 12 Dec 2025 19:24:26 +0800</pubDate>
      <guid>https://www.neican.ai/newspaper/2025-12-12-ai-agent-token-/</guid>
      <description>今天是2025年12月12日。智能体（AI Agent）正以势不可挡的姿态，全面重构我们对技术、商业乃至人类自身边界的认知。这场从“Token内卷”到“星辰大海”的“Agent战争”，正深刻重塑AI行业的底层逻辑与未来走向。</description>
    </item>
    <item>
      <title>颠覆“逐字生成”：蚂蚁集团LLaDA 2.0开启AI可控生成新纪元</title>
      <link>https://www.neican.ai/insights/llada-20ai-20251212154008175-2/</link>
      <pubDate>Fri, 12 Dec 2025 15:40:08 +0800</pubDate>
      <guid>https://www.neican.ai/insights/llada-20ai-20251212154008175-2/</guid>
      <description>蚂蚁集团携手多方力量，推出千亿参数级扩散语言模型LLaDA 2.0，其核心在于改变了传统自回归模型的“逐字生成”模式，实现推理过程中Token的直接编辑与控制。这一颠覆性创新不仅有望显著提升AI生成效率和降低成本，更重要的是，它为AI内容带来了前所未有的可控性和可编辑性，预示着AI在编程、创意写作等领域的人机协作将迎来范式级的深刻变革。</description>
    </item>
    <item>
      <title>超越自回归：蚂蚁dInfer如何将扩散语言模型推向AGI主干道</title>
      <link>https://www.neican.ai/insights/dinferagi-20251013171004788-0/</link>
      <pubDate>Mon, 13 Oct 2025 17:10:04 +0800</pubDate>
      <guid>https://www.neican.ai/insights/dinferagi-20251013171004788-0/</guid>
      <description>蚂蚁集团开源dInfer推理框架，显著提升扩散语言模型（dLLM）推理速度超10倍，使其在效率上首次超越自回归模型，结合其LLaDA-MoE模型在精度上的突破，共同为通用人工智能（AGI）的实现开辟了一条高效且具竞争力的“主干道”。这一进展不仅大幅降低了AI应用的商业化成本，还将重塑AI架构范式，加速产业智能化进程并引发深远的社会影响。</description>
    </item>
    <item>
      <title>轨迹感知RL突破扩散模型瓶颈：AI推理范式的新统一与效率革命</title>
      <link>https://www.neican.ai/insights/rlai-20250915181004754-1/</link>
      <pubDate>Mon, 15 Sep 2025 18:10:04 +0800</pubDate>
      <guid>https://www.neican.ai/insights/rlai-20250915181004754-1/</guid>
      <description>TraceRL框架通过创新性地对齐扩散语言模型（DLM）的训练目标与推理轨迹，显著提升了DLM的性能与训练效率，使其在复杂推理任务上超越了更大规模的自回归模型，预示着AI推理能力与计算效率的深刻变革。这一创新及其开源框架dLLM-RL，正加速DLM成为下一代高效智能体的核心基石，开启“RL大一统”的新范式。</description>
    </item>
    <item>
      <title>LLaDA-MoE：扩散语言模型的MoE化挑战自回归霸权，效率与精度的新范式之争</title>
      <link>https://www.neican.ai/insights/llada-moemoe-20250912111004898-5/</link>
      <pubDate>Fri, 12 Sep 2025 11:10:04 +0800</pubDate>
      <guid>https://www.neican.ai/insights/llada-moemoe-20250912111004898-5/</guid>
      <description>蚂蚁集团与人民大学联合发布的LLaDA-MoE，是全球首个原生MoE架构的扩散语言模型，其性能与主流自回归模型Qwen2.5-3B-Instruct持平，并在推理速度上占据优势，为AGI探索开辟了新路径。然而，该技术仍需在应对“效率悖论”和提升序列级准确性方面持续优化，以在更广泛的商业应用中证明其价值。</description>
    </item>
  </channel>
</rss>
