LLM的顿悟现象

✨步子哥 (steper) • 2025年12月22日 07:12

                        <!DOCTYPE html><html lang="zh-CN"><head>
    <meta charset="UTF-8"/>
    <meta name="viewport" content="width=device-width, initial-scale=1.0"/>
    <title>归纳偏置：解锁Grokking与模型泛化之谜的钥匙</title>
    <script src="https://cdn.tailwindcss.com"></script>
    <link href="https://fonts.googleapis.com/css2?family=Inter:wght@300;400;500;600;700&amp;family=Noto+Sans+SC:wght@300;400;500;600;700&amp;family=Noto+Serif+SC:wght@400;600;700&amp;display=swap" rel="stylesheet"/>
    <link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/6.4.0/css/all.min.css"/>
    <script src="https://cdn.jsdelivr.net/npm/mermaid@10.6.1/dist/mermaid.min.js"></script>
    <style>
        :root {
            --primary: #2563eb;
            --secondary: #64748b;
            --accent: #f59e0b;
            --background: #f8fafc;
            --surface: #ffffff;
            --text: #1e293b;
            --text-muted: #64748b;
            --border: #e2e8f0;
            --success: #10b981;
            --warning: #f59e0b;
            --error: #ef4444;
        }
        
        body {
            font-family: 'Inter', 'Noto Sans SC', sans-serif;
            background: var(--background);
            color: var(--text);
            line-height: 1.7;
            max-width: 100vw;
            overflow-x: hidden;
        }
        
        .serif {
            font-family: 'Noto Serif SC', serif;
        }
        
        .toc-sidebar {
            position: fixed;
            left: 0;
            top: 0;
            width: 280px;
            height: 100vh;
            background: var(--surface);
            border-right: 1px solid var(--border);
            overflow-y: auto;
            z-index: 1000;
            padding: 2rem 1.5rem;
        }
        
        .main-content {
            margin: 10px;
            min-height: 100vh;
        }
        
        .hero-grid {
            display: grid;
            grid-template-columns: 1fr 1fr;
            gap: 2rem;
            align-items: center;
            min-height: 60vh;
        }
        
        .hero-title {
            font-size: 3.5rem;
            font-weight: 700;
            line-height: 1.1;
            background: linear-gradient(135deg, var(--primary), var(--accent));
            -webkit-background-clip: text;
            -webkit-text-fill-color: transparent;
            background-clip: text;
        }
        
        .hero-subtitle {
            font-size: 1.25rem;
            color: var(--text-muted);
            margin-top: 1rem;
        }
        
        .citation {
            color: var(--primary);
            text-decoration: none;
            font-weight: 500;
            border-bottom: 1px dotted var(--primary);
        }
        
        .citation:hover {
            background: rgba(37, 99, 235, 0.1);
            padding: 0 2px;
            border-radius: 2px;
        }
        
        .section-divider {
            height: 1px;
            background: linear-gradient(90deg, transparent, var(--border), transparent);
            margin: 4rem 0;
        }
        
        .highlight-box {
            background: linear-gradient(135deg, rgba(37, 99, 235, 0.05), rgba(245, 158, 11, 0.05));
            border-left: 4px solid var(--primary);
            padding: 1.5rem;
            margin: 2rem 0;
            border-radius: 0 8px 8px 0;
        }
        
        .quote-block {
            font-style: italic;
            font-size: 1.125rem;
            color: var(--text-muted);
            border-left: 3px solid var(--accent);
            padding-left: 1.5rem;
            margin: 2rem 0;
        }
        
        .mermaid-container {
            display: flex;
            justify-content: center;
            min-height: 300px;
            max-height: 800px;
            background: #ffffff;
            border: 2px solid #e5e7eb;
            border-radius: 12px;
            padding: 30px;
            margin: 30px 0;
            box-shadow: 0 8px 25px rgba(0, 0, 0, 0.08);
            position: relative;
            overflow: hidden;
        }

        .mermaid-container .mermaid {
            width: 100%;
            max-width: 100%;
            height: 100%;
            cursor: grab;
            transition: transform 0.3s ease;
            transform-origin: center center;
            display: flex;
            justify-content: center;
            align-items: center;
            touch-action: none;
            -webkit-user-select: none;
            -moz-user-select: none;
            -ms-user-select: none;
            user-select: none;
        }

        .mermaid-container .mermaid svg {
            max-width: 100%;
            height: 100%;
            display: block;
            margin: 0 auto;
        }

        .mermaid-container .mermaid:active {
            cursor: grabbing;
        }

        .mermaid-container.zoomed .mermaid {
            height: 100%;
            width: 100%;
            cursor: grab;
        }

        .mermaid-controls {
            position: absolute;
            top: 15px;
            right: 15px;
            display: flex;
            gap: 10px;
            z-index: 20;
            background: rgba(255, 255, 255, 0.95);
            padding: 8px;
            border-radius: 8px;
            box-shadow: 0 2px 8px rgba(0, 0, 0, 0.1);
        }

        .mermaid-control-btn {
            background: #ffffff;
            border: 1px solid #d1d5db;
            border-radius: 6px;
            padding: 10px;
            cursor: pointer;
            transition: all 0.2s ease;
            color: #374151;
            font-size: 14px;
            min-width: 36px;
            height: 36px;
            text-align: center;
            display: flex;
            align-items: center;
            justify-content: center;
        }

        .mermaid-control-btn:hover {
            background: #f8fafc;
            border-color: #3b82f6;
            color: #3b82f6;
            transform: translateY(-1px);
        }

        .mermaid-control-btn:active {
            transform: scale(0.95);
        }
        
        .toc-link {
            display: block;
            padding: 0.5rem 0;
            color: var(--text-muted);
            text-decoration: none;
            border-bottom: 1px solid transparent;
            transition: all 0.2s ease;
        }
        
        .toc-link:hover {
            color: var(--primary);
            border-bottom-color: var(--primary);
        }
        
        .toc-link.active {
            color: var(--primary);
            font-weight: 500;
        }
        
        .section-number {
            color: var(--accent);
            font-weight: 600;
            margin-right: 0.5rem;
        }
        
        <span class="mention-invalid">@media</span> (max-width: 1024px) {
            .toc-sidebar {
                transform: translateX(-100%);
                transition: transform 0.3s ease;
            }
            
            .toc-sidebar.open {
                transform: translateX(0);
            }
            
            .main-content {
                margin-left: 0;
            }
            
            .hero-grid {
                grid-template-columns: 1fr;
                gap: 1rem;
            }
            
            .hero-title {
                font-size: 2.5rem;
            }

            .mermaid-control-btn:not(.reset-zoom) {
                display: none;
            }
            .mermaid-controls {
                top: auto;
                bottom: 15px;
                right: 15px;
            }
        }
        
        <span class="mention-invalid">@media</span> (max-width: 768px) {
            .hero-title {
                font-size: 2rem;
            }
            .hero-subtitle {
                font-size: 1rem;
            }
            .px-8 {
                padding-left: 1rem !important;
                padding-right: 1rem !important;
            }
        }
    </style>
  <base target="_blank">
</head>

  <body>
    <!-- Main Content -->
    <main class="main-content">
      
      <!-- Introduction Section -->
      <section id="introduction" class="py-16 bg-white">
        <div class="container mx-auto px-8 max-w-4xl">
          <div class="highlight-box">
            <h2 class="text-2xl font-bold mb-4 serif">核心洞察</h2>
            <p class="text-lg leading-relaxed">
              归纳偏置是驱动Grokking现象的核心机制。它通过一个在训练过程中动态演化的&#34;相变&#34;过程来实现：早期，模型的偏置（如大初始化带来的隐式偏置）倾向于快速拟合训练数据的&#34;记忆解&#34;；晚期，另一种偏置（如权重衰减的显式偏置或Adam优化器的&#34;Slingshot&#34;隐式偏置）占据主导，将模型推向更简洁、更具泛化能力的&#34;泛化解&#34;。
            </p>
          </div>

          <div class="grid md:grid-cols-2 gap-8 mt-12">
            <div class="space-y-4">
              <h3 class="text-xl font-semibold serif">理论基础</h3>
              <p>
                这一从记忆到泛化的转变，在宏观上表现为模型性能的&#34;顿悟&#34;式提升。该过程可通过优化动态的二分性、神经网络内部&#34;记忆电路&#34;与&#34;泛化电路&#34;的竞争，以及对模型复杂度的系统性偏好等多个层面进行解释。
              </p>
            </div>
            <div class="space-y-4">
              <h3 class="text-xl font-semibold serif">研究范围</h3>
              <p>
                近期的研究，特别是从2022年至2025年的系列工作，将焦点指向了归纳偏置这一核心驱动力。这些研究揭示了Grokking并非偶然，而是深度学习优化过程中内在偏置与数据结构相互作用的必然结果。
              </p>
            </div>
          </div>
        </div>
      </section>

      <div class="section-divider"></div>

      <!-- Mechanisms Section -->
      <section id="mechanisms" class="py-16 bg-slate-50">
        <div class="container mx-auto px-8 max-w-6xl">
          <h2 class="text-3xl font-bold mb-12 serif text-center">核心机制：多层次解释框架</h2>

          <!-- Dichotomy Subsection -->
          <div id="dichotomy" class="mb-16">
            <h3 class="text-2xl font-semibold mb-8 serif">优化动态的二分性：从记忆到泛化的转变</h3>

            <div class="grid lg:grid-cols-3 gap-8 mb-8">
              <div class="bg-white p-6 rounded-xl shadow-lg">
                <div class="w-12 h-12 bg-blue-100 rounded-lg flex items-center justify-center mb-4">
                  <i class="fas fa-clock text-blue-600"></i>
                </div>
                <h4 class="font-semibold mb-3">早期阶段</h4>
                <p class="text-sm text-gray-600">大初始化带来的隐式偏置，使模型进入&#34;懒惰&#34;训练模式，快速拟合训练数据</p>
              </div>
              <div class="bg-white p-6 rounded-xl shadow-lg">
                <div class="w-12 h-12 bg-green-100 rounded-lg flex items-center justify-center mb-4">
                  <i class="fas fa-arrows-alt text-green-600"></i>
                </div>
                <h4 class="font-semibold mb-3">相变点</h4>
                <p class="text-sm text-gray-600">权重衰减的累积效应开始主导，推动模型向最小范数解转变</p>
              </div>
              <div class="bg-white p-6 rounded-xl shadow-lg">
                <div class="w-12 h-12 bg-purple-100 rounded-lg flex items-center justify-center mb-4">
                  <i class="fas fa-lightbulb text-purple-600"></i>
                </div>
                <h4 class="font-semibold mb-3">晚期阶段</h4>
                <p class="text-sm text-gray-600">模型切换到&#34;丰富&#34;模式，学习有意义的特征表示，实现泛化</p>
              </div>
            </div>

            <div class="quote-block">
              &#34;Grokking的相变可以被理解为：在训练的漫长过程中，权重衰减的&#39;拉力&#39;最终克服了早期隐式偏置的&#39;推力&#39;，将模型从一个记忆解的吸引子盆地中&#39;拽&#39;出。&#34;
            </div>

            <div class="bg-white p-6 rounded-xl shadow-lg">
              <h4 class="font-semibold mb-4">理论支持：Lyu等人的阶段二分理论证明</h4>
              <p class="text-gray-700 mb-4">
                <a href="#" class="citation">Lyu等人（2023-2024）</a>的研究工作为Grokking现象提供了严谨的数学证明。他们的理论分析表明，优化过程可以被清晰地划分为两个不同阶段：
              </p>
              <ul class="list-disc list-inside space-y-2 text-gray-600">
                <li><strong>第一阶段：</strong>模型动态主要由初始化决定，行为类似核方法，目标是完美拟合训练数据</li>
                <li><strong>第二阶段：</strong>权重衰减开始主导，引导优化器寻找最小范数解，实现泛化</li>
              </ul>
            </div>
          </div>

          <!-- Circuit Competition -->
          <div id="circuit-competition" class="mb-16">
            <h3 class="text-2xl font-semibold mb-8 serif">表示学习视角：电路竞争与效率偏好</h3>

            <div class="grid lg:grid-cols-2 gap-8 mb-8">
              <div class="bg-white p-6 rounded-xl shadow-lg">
                <h4 class="font-semibold mb-4 text-red-600">记忆电路</h4>
                <ul class="space-y-2 text-gray-600">
                  <li>• 结构复杂、参数冗余</li>
                  <li>• 快速实现零误差拟合</li>
                  <li>• 存储特定样本映射</li>
                  <li>• 泛化能力差</li>
                </ul>
              </div>
              <div class="bg-white p-6 rounded-xl shadow-lg">
                <h4 class="font-semibold mb-4 text-green-600">泛化电路</h4>
                <ul class="space-y-2 text-gray-600">
                  <li>• 结构简洁、计算高效</li>
                  <li>• 学习潜在规律</li>
                  <li>• 捕捉可泛化模式</li>
                  <li>• 优秀的泛化能力</li>
                </ul>
              </div>
            </div>

            <div class="highlight-box">
              <h4 class="font-semibold mb-3">权重衰减对高效电路的偏好</h4>
              <p class="mb-4">
                权重衰减不仅惩罚参数的大范数，更深层次地偏好那些具有<strong>低秩</strong>或<strong>稀疏</strong>结构的表示。这种偏置可以被看作是<strong>秩最小化</strong>的代理。
              </p>
              <p>
                当低秩的泛化电路在效率上最终超越高秩的记忆电路时，Grokking的相变就发生了。这个临界点可以被精确地定义为泛化电路的效率首次超过记忆电路效率的时刻。
              </p>
            </div>
          </div>

          <!-- Optimizer Bias -->
          <div id="optimizer" class="mb-16">
            <h3 class="text-2xl font-semibold mb-8 serif">优化器特定偏置：Adam的&#34;Slingshot&#34;机制</h3>

            <div class="bg-white p-6 rounded-xl shadow-lg">
              <div class="flex items-center mb-4">
                <div class="w-16 h-16 bg-gradient-to-br from-orange-400 to-red-500 rounded-xl flex items-center justify-center mr-4">
                  <i class="fas fa-rocket text-white text-xl"></i>
                </div>
                <div>
                  <h4 class="font-semibold text-lg">Slingshot机制</h4>
                  <p class="text-gray-600">自适应优化器晚期的异常动态，意外引入泛化偏置</p>
                </div>
              </div>

              <p class="text-gray-700 mb-4">
                <a href="#" class="citation">Thilak等人（2022）</a>的研究发现，Adam优化器在训练后期可能出现一种周期性的、非单调的动态，即&#34;Slingshot&#34;效应。这种看似不稳定的动态，实际上是一种有益的隐式偏置。
              </p>

              <div class="grid md:grid-cols-3 gap-4 mt-6">
                <div class="text-center p-4 bg-orange-50 rounded-lg">
                  <i class="fas fa-compress-arrows-alt text-orange-600 text-2xl mb-2"></i>
                  <p class="text-sm font-medium">暂时陷入局部最小值</p>
                </div>
                <div class="text-center p-4 bg-blue-50 rounded-lg">
                  <i class="fas fa-chart-line text-blue-600 text-2xl mb-2"></i>
                  <p class="text-sm font-medium">积累能量准备挣脱</p>
                </div>
                <div class="text-center p-4 bg-green-50 rounded-lg">
                  <i class="fas fa-paper-plane text-green-600 text-2xl mb-2"></i>
                  <p class="text-sm font-medium">弹射到泛化区域</p>
                </div>
              </div>
            </div>
          </div>

          <!-- Complexity Bias -->
          <div id="complexity" class="mb-16">
            <h3 class="text-2xl font-semibold mb-8 serif">复杂度与秩最小化偏置</h3>

            <div class="grid lg:grid-cols-2 gap-8">
              <div class="bg-white p-6 rounded-xl shadow-lg">
                <h4 class="font-semibold mb-4">从高秩记忆解到低秩泛化解</h4>
                <p class="text-gray-700 mb-4">
                  Grokking的核心转变可以表述为从高秩的记忆表示到低秩的泛化表示的跃迁。这个转变点对应于模型内部关键权重矩阵的秩发生显著下降的时刻。
                </p>
                <div class="flex items-center justify-between mt-6 p-4 bg-gray-50 rounded-lg">
                  <span class="text-sm font-medium">初始秩</span>
                  <span class="text-2xl font-bold text-red-600">高</span>
                  <i class="fas fa-arrow-right text-gray-400"></i>
                  <span class="text-sm font-medium">最终秩</span>
                  <span class="text-2xl font-bold text-green-600">低</span>
                </div>
              </div>

              <div class="bg-white p-6 rounded-xl shadow-lg">
                <h4 class="font-semibold mb-4">任务特定复杂度偏置的必要性</h4>
                <p class="text-gray-700 mb-4">
                  2025年最新研究表明，简单性偏置并非万能。在某些复杂的表格数据、高维回归问题中，过于强烈的简单性偏置可能是有害的。
                </p>
                <div class="bg-yellow-50 p-4 rounded-lg mt-4">
                  <p class="text-sm text-yellow-800">
                    <i class="fas fa-exclamation-triangle mr-2"></i>
                    理想的归纳偏置应该与任务的真实复杂度相匹配，而不是一味地追求简单。
                  </p>
                </div>
              </div>
            </div>
          </div>
        </div>
      </section>

      <div class="section-divider"></div>

      <!-- Empirical Observations -->
      <section id="empirical" class="py-16 bg-white">
        <div class="container mx-auto px-8 max-w-6xl">
          <h2 class="text-3xl font-bold mb-12 serif text-center">实证观察：归纳偏置在不同场景下的表现</h2>

          <!-- Optimizer and Regularization -->
          <div id="optimizer-regularization" class="mb-16">
            <h3 class="text-2xl font-semibold mb-8 serif">优化器与正则化方法的影响</h3>

            <div class="grid lg:grid-cols-3 gap-8 mb-8">
              <div class="bg-gradient-to-br from-blue-50 to-blue-100 p-6 rounded-xl">
                <h4 class="font-semibold mb-4 text-blue-800">Adam优化器</h4>
                <ul class="space-y-2 text-sm text-blue-700">
                  <li>• Slingshot机制</li>
                  <li>• 非单调动态</li>
                  <li>• 更强的探索能力</li>
                  <li>• 更容易触发Grokking</li>
                </ul>
              </div>
              <div class="bg-gradient-to-br from-gray-50 to-gray-100 p-6 rounded-xl">
                <h4 class="font-semibold mb-4 text-gray-800">SGD优化器</h4>
                <ul class="space-y-2 text-sm text-gray-700">
                  <li>• 平滑收敛轨迹</li>
                  <li>• 稳定局部最小值</li>
                  <li>• 较弱探索能力</li>
                  <li>• 较难触发Grokking</li>
                </ul>
              </div>
              <div class="bg-gradient-to-br from-green-50 to-green-100 p-6 rounded-xl">
                <h4 class="font-semibold mb-4 text-green-800">权重衰减</h4>
                <ul class="space-y-2 text-sm text-green-700">
                  <li>• 核心显式偏置</li>
                  <li>• 强度调节相变</li>
                  <li>• 惩罚大范数参数</li>
                  <li>• 引导最小范数解</li>
                </ul>
              </div>
            </div>

            <div class="bg-white p-6 rounded-xl shadow-lg border-l-4 border-orange-400">
              <h4 class="font-semibold mb-4">权重衰减强度的影响</h4>
              <div class="grid md:grid-cols-3 gap-4">
                <div class="text-center">
                  <div class="w-16 h-16 bg-red-100 rounded-full flex items-center justify-center mx-auto mb-2">
                    <i class="fas fa-times text-red-600"></i>
                  </div>
                  <p class="font-medium text-red-600">过低或无</p>
                  <p class="text-sm text-gray-600">模型倾向于记忆，Grokking不发生</p>
                </div>
                <div class="text-center">
                  <div class="w-16 h-16 bg-green-100 rounded-full flex items-center justify-center mx-auto mb-2">
                    <i class="fas fa-check text-green-600"></i>
                  </div>
                  <p class="font-medium text-green-600">适中强度</p>
                  <p class="text-sm text-gray-600">理想情况，清晰的相变</p>
                </div>
                <div class="text-center">
                  <div class="w-16 h-16 bg-orange-100 rounded-full flex items-center justify-center mx-auto mb-2">
                    <i class="fas fa-exclamation text-orange-600"></i>
                  </div>
                  <p class="font-medium text-orange-600">过高强度</p>
                  <p class="text-sm text-gray-600">过度惩罚，欠拟合</p>
                </div>
              </div>
            </div>
          </div>

          <!-- Architecture and Task Types -->
          <div id="architecture" class="mb-16">
            <h3 class="text-2xl font-semibold mb-8 serif">模型架构与任务类型的影响</h3>

            <div class="space-y-8">
              <div class="bg-white p-6 rounded-xl shadow-lg">
                <h4 class="font-semibold mb-4 text-blue-600">
                  <i class="fas fa-calculator mr-2"></i>玩具任务（模加法）
                </h4>
                <p class="text-gray-700 mb-4">
                  在模加法 `a + b ≡ c (mod p)` 等算法任务中，由于内在规律简洁，泛化电路效率远高于记忆电路，Grokking的相变通常表现得<strong>清晰、尖锐且可重复</strong>。
                </p>
                <div class="bg-blue-50 p-4 rounded-lg">
                  <p class="text-sm text-blue-800">
                    <i class="fas fa-info-circle mr-2"></i>
                    这些任务为验证Grokking的理论模型提供了理想的实验环境
                  </p>
                </div>
              </div>

              <div class="bg-white p-6 rounded-xl shadow-lg">
                <h4 class="font-semibold mb-4 text-purple-600">
                  <i class="fas fa-database mr-2"></i>表格数据与回归任务
                </h4>
                <p class="text-gray-700 mb-4">
                  真实世界的数据通常不包含单一完美的数学规律，其内在规律可能更加复杂、嘈杂。在这种情况下，简单性偏置的效果变得有限，模型可能需要学习<strong>更高复杂度的函数</strong>。
                </p>
                <div class="bg-purple-50 p-4 rounded-lg">
                  <p class="text-sm text-purple-800">
                    <i class="fas fa-lightbulb mr-2"></i>
                    这揭示了归纳偏置的任务依赖性，不存在&#34;放之四海而皆准&#34;的最佳偏置
                  </p>
                </div>
              </div>

              <div class="bg-white p-6 rounded-xl shadow-lg">
                <h4 class="font-semibold mb-4 text-green-600">
                  <i class="fas fa-language mr-2"></i>大语言模型（LLM）
                </h4>
                <p class="text-gray-700 mb-4">
                  在LLM中，Grokking呈现出<strong>局部与异步</strong>的特征。模型的不同能力（语法、事实知识、推理）可能在不同时间点、基于不同数据子集发生Grokking。
                </p>
                <div class="bg-green-50 p-4 rounded-lg">
                  <p class="text-sm text-green-800">
                    <i class="fas fa-network-wired mr-2"></i>
                    LLM的&#34;涌现能力&#34;可视为一系列在不同领域、不同时间上发生的局部Grokking事件的集合
                  </p>
                </div>
              </div>
            </div>
          </div>

          <!-- Limitations -->
          <div id="limitations" class="mb-16">
            <h3 class="text-2xl font-semibold mb-8 serif">归纳偏置的负面效应与局限性</h3>

            <div class="grid lg:grid-cols-3 gap-8">
              <div class="bg-red-50 p-6 rounded-xl border-l-4 border-red-400">
                <h4 class="font-semibold mb-4 text-red-800">简单性偏置的局限</h4>
                <p class="text-sm text-red-700">
                  在复杂任务中，强制模型学习简单函数的偏置会导致无法充分拟合数据，限制性能上限，起到&#34;矫枉过正&#34;的作用。
                </p>
              </div>

              <div class="bg-orange-50 p-6 rounded-xl border-l-4 border-orange-400">
                <h4 class="font-semibold mb-4 text-orange-800">任务依赖性</h4>
                <p class="text-sm text-orange-700">
                  最佳偏置高度依赖于任务，缺乏统一的理论指导如何选择最优偏置，这使得理论难以直接推广。
                </p>
              </div>

              <div class="bg-yellow-50 p-6 rounded-xl border-l-4 border-yellow-400">
                <h4 class="font-semibold mb-4 text-yellow-800">计算成本</h4>
                <p class="text-sm text-yellow-700">
                  Grokking通常需要超长训练，带来巨大的计算成本。在实际应用中，训练直到Grokking可能不经济。
                </p>
              </div>
            </div>
          </div>
        </div>
      </section>

      <div class="section-divider"></div>

      <!-- Applications Section -->
      <section id="applications" class="py-16 bg-slate-50">
        <div class="container mx-auto px-8 max-w-6xl">
          <h2 class="text-3xl font-bold mb-12 serif text-center">应用与启示：利用归纳偏置设计更好的模型</h2>

          <!-- Training Strategies -->
          <div id="training-strategies" class="mb-16">
            <h3 class="text-2xl font-semibold mb-8 serif">加速与控制Grokking的训练策略</h3>

            <div class="grid lg:grid-cols-2 gap-8 mb-8">
              <div class="bg-white p-6 rounded-xl shadow-lg">
                <h4 class="font-semibold mb-4 text-blue-600">调整权重衰减与初始化</h4>
                <div class="space-y-4">
                  <div class="bg-blue-50 p-4 rounded-lg">
                    <h5 class="font-medium text-blue-800 mb-2">课程式权重衰减</h5>
                    <p class="text-sm text-blue-700">初期使用较小权重衰减，后期逐渐增加强度，加速相变</p>
                  </div>
                  <div class="bg-blue-50 p-4 rounded-lg">
                    <h5 class="font-medium text-blue-800 mb-2">多尺度初始化</h5>
                    <p class="text-sm text-blue-700">从大初始化开始确保&#34;懒惰&#34;模式，为后续&#34;顿悟&#34;创造条件</p>
                  </div>
                </div>
              </div>

              <div class="bg-white p-6 rounded-xl shadow-lg">
                <h4 class="font-semibold mb-4 text-green-600">优化器选择</h4>
                <div class="space-y-4">
                  <div class="bg-green-50 p-4 rounded-lg">
                    <h5 class="font-medium text-green-800 mb-2">优先选择Adam</h5>
                    <p class="text-sm text-green-700">自适应优化器的Slingshot机制天然适合诱导Grokking</p>
                  </div>
                  <div class="bg-green-50 p-4 rounded-lg">
                    <h5 class="font-medium text-green-800 mb-2">增强隐式偏置</h5>
                    <p class="text-sm text-green-700">设计特定动量策略或学习率调度，增强泛化偏置</p>
                  </div>
                </div>
              </div>
            </div>

            <div class="highlight-box">
              <h4 class="font-semibold mb-4">监控电路/秩演化作为Grokking指标</h4>
              <p class="mb-4">
                通过定期计算关键权重矩阵的秩或奇异值分布，可以预测Grokking的相变。当观察到秩开始显著下降时，通常预示着相变即将发生。
              </p>
              <div class="flex items-center space-x-4 mt-4">
                <div class="flex-1 bg-gradient-to-r from-blue-200 to-green-200 h-2 rounded-full">
                  <div class="bg-gradient-to-r from-blue-500 to-green-500 h-2 rounded-full" style="width: 75%"></div>
                </div>
                <span class="text-sm text-gray-600">秩下降进度</span>
              </div>
            </div>
          </div>

          <!-- LLM Abilities -->
          <div id="llm-abilities" class="mb-16">
            <h3 class="text-2xl font-semibold mb-8 serif">理解大语言模型的涌现能力</h3>

            <div class="bg-white p-8 rounded-xl shadow-lg">
              <div class="grid lg:grid-cols-2 gap-8">
                <div>
                  <h4 class="font-semibold mb-4 text-purple-600">局部Grokking与特定能力的关联</h4>
                  <p class="text-gray-700 mb-4">
                    LLM的每种高级能力都可以看作独立的&#34;任务&#34;，经历从记忆到泛化的Grokking过程：
                  </p>
                  <div class="space-y-3">
                    <div class="flex items-center space-x-3">
                      <div class="w-3 h-3 bg-blue-500 rounded-full"></div>
                      <span class="text-sm"><strong>语法能力：</strong>学习句法结构的规律</span>
                    </div>
                    <div class="flex items-center space-x-3">
                      <div class="w-3 h-3 bg-green-500 rounded-full"></div>
                      <span class="text-sm"><strong>事实知识：</strong>记忆和关联世界知识</span>
                    </div>
                    <div class="flex items-center space-x-3">
                      <div class="w-3 h-3 bg-purple-500 rounded-full"></div>
                      <span class="text-sm"><strong>推理能力：</strong>进行多步逻辑推断</span>
                    </div>
                  </div>
                </div>

                <div>
                  <img src="https://kimi-web-img.moonshot.cn/img/pic3.zhimg.com/c926862e1805cfe88d9cb9252d9505ed21e17005.jpg" alt="大语言模型能力涌现的抽象可视化" class="w-full rounded-lg shadow-md" size="medium" aspect="wide" query="大语言模型涌现能力抽象表示" referrerpolicy="no-referrer" data-modified="1" data-score="0.00"/>
                </div>
              </div>

              <div class="mt-8 p-6 bg-purple-50 rounded-lg">
                <h5 class="font-semibold text-purple-800 mb-3">归纳偏置作为涌现能力的根本驱动力</h5>
                <p class="text-purple-700 text-sm">
                  Transformer架构（注意力机制、位置编码）和训练目标（预测下一个词）共同引入强大的结构偏置，引导模型在海量数据中寻找有意义的模式。当积累达到临界点时，Grokking发生，新能力涌现。
                </p>
              </div>
            </div>
          </div>

          <!-- Task-Specific Bias -->
          <div id="task-specific" class="mb-16">
            <h3 class="text-2xl font-semibold mb-8 serif">设计任务特定的归纳偏置</h3>

            <div class="grid lg:grid-cols-3 gap-8">
              <div class="bg-white p-6 rounded-xl shadow-lg">
                <div class="w-12 h-12 bg-blue-100 rounded-lg flex items-center justify-center mb-4">
                  <i class="fas fa-layer-group text-blue-600"></i>
                </div>
                <h4 class="font-semibold mb-3">结构化正则化</h4>
                <p class="text-sm text-gray-600 mb-4">设计能够鼓励特定网络结构的正则化项，如层级化、模块化</p>
                <div class="bg-blue-50 p-3 rounded-lg">
                  <p class="text-xs text-blue-700">促进模块化表示学习</p>
                </div>
              </div>

              <div class="bg-white p-6 rounded-xl shadow-lg">
                <div class="w-12 h-12 bg-green-100 rounded-lg flex items-center justify-center mb-4">
                  <i class="fas fa-compress-arrows-alt text-green-600"></i>
                </div>
                <h4 class="font-semibold mb-3">信息论偏置</h4>
                <p class="text-sm text-gray-600 mb-4">基于信息瓶颈或最小描述长度原理的偏置</p>
                <div class="bg-green-50 p-3 rounded-lg">
                  <p class="text-xs text-green-700">学习紧凑且信息丰富的表示</p>
                </div>
              </div>

              <div class="bg-white p-6 rounded-xl shadow-lg">
                <div class="w-12 h-12 bg-purple-100 rounded-lg flex items-center justify-center mb-4">
                  <i class="fas fa-dna text-purple-600"></i>
                </div>
                <h4 class="font-semibold mb-3">因果偏置</h4>
                <p class="text-sm text-gray-600 mb-4">在模型中引入对因果关系的偏好</p>
                <div class="bg-purple-50 p-3 rounded-lg">
                  <p class="text-xs text-purple-700">学习更鲁棒的规律</p>
                </div>
              </div>
            </div>

            <div class="mt-8 bg-white p-6 rounded-xl shadow-lg">
              <h4 class="font-semibold mb-4 text-center">平衡泛化能力与模型容量</h4>
              <p class="text-center text-gray-600 mb-6">
                设计任务特定的归纳偏置是在泛化能力和模型容量之间进行权衡的艺术
              </p>
              <div class="flex items-center justify-center space-x-8">
                <div class="text-center">
                  <div class="w-16 h-16 bg-blue-100 rounded-full flex items-center justify-center mb-2">
                    <i class="fas fa-brain text-blue-600 text-xl"></i>
                  </div>
                  <p class="font-medium">泛化能力</p>
                </div>
                <div class="text-4xl text-gray-300">⚖️</div>
                <div class="text-center">
                  <div class="w-16 h-16 bg-orange-100 rounded-full flex items-center justify-center mb-2">
                    <i class="fas fa-expand-arrows-alt text-orange-600 text-xl"></i>
                  </div>
                  <p class="font-medium">模型容量</p>
                </div>
              </div>
            </div>
          </div>
        </div>
      </section>

      <div class="section-divider"></div>

      <!-- Future Directions -->
      <section id="future" class="py-16 bg-white">
        <div class="container mx-auto px-8 max-w-6xl">
          <h2 class="text-3xl font-bold mb-12 serif text-center">未来方向与开放问题</h2>

          <!-- Theoretical Mysteries -->
          <div id="theoretical" class="mb-16">
            <h3 class="text-2xl font-semibold mb-8 serif">理论层面的未解之谜</h3>

            <div class="grid lg:grid-cols-2 gap-8">
              <div class="space-y-6">
                <div class="bg-white p-6 rounded-xl shadow-lg border-l-4 border-blue-400">
                  <h4 class="font-semibold mb-3 text-blue-800">统一Grokking理论的构建</h4>
                  <p class="text-gray-700 mb-4">
                    当前关于Grokking的解释多样但零散，缺乏统一的数学框架。未来的关键挑战是构建一个能够整合所有视角的统一理论。
                  </p>
                  <div class="bg-blue-50 p-3 rounded-lg">
                    <p class="text-sm text-blue-700">
                      需要解释为什么不同条件下不同机制会占据主导，以及这些机制如何相互作用
                    </p>
                  </div>
                </div>

                <div class="bg-white p-6 rounded-xl shadow-lg border-l-4 border-green-400">
                  <h4 class="font-semibold mb-3 text-green-800">无正则化下纯隐式偏置的极限</h4>
                  <p class="text-gray-700 mb-4">
                    目前理论多依赖显式正则化。理解纯隐式偏置（仅由优化器动态驱动）的极限是一个重要的开放问题。
                  </p>
                  <div class="bg-green-50 p-3 rounded-lg">
                    <p class="text-sm text-green-700">
                      需要回答：在没有权重衰减的情况下，优化器的隐式偏置是否足以驱动泛化？
                    </p>
                  </div>
                </div>
              </div>

              <div class="space-y-6">
                <div class="bg-white p-6 rounded-xl shadow-lg border-l-4 border-purple-400">
                  <h4 class="font-semibold mb-3 text-purple-800">归纳偏置与泛化边界的精确关系</h4>
                  <p class="text-gray-700 mb-4">
                    虽然知道归纳偏置能促进泛化，但对其与泛化边界的精确关系知之甚少。
                  </p>
                  <div class="bg-purple-50 p-3 rounded-lg">
                    <p class="text-sm text-purple-700">
                      需要探索如何将偏置直接纳入泛化边界的推导中
                    </p>
                  </div>
                </div>

                <div class="bg-gray-50 p-6 rounded-xl">
                  <h4 class="font-semibold mb-3 text-gray-800">开放问题总结</h4>
                  <ul class="space-y-2 text-sm text-gray-700">
                    <li>• 如何设计任务特定归纳偏置，避免简单性偏置局限？</li>
                    <li>• LLM涌现能力是否主要源于类似偏置相变？</li>
                    <li>• 无正则化下Grokking的纯隐式偏置极限何在？</li>
                  </ul>
                </div>
              </div>
            </div>
          </div>

          <!-- Practical Challenges -->
          <div id="practical" class="mb-16">
            <h3 class="text-2xl font-semibold mb-8 serif">实践层面的挑战与机遇</h3>

            <div class="grid lg:grid-cols-3 gap-8">
              <div class="bg-white p-6 rounded-xl shadow-lg">
                <div class="w-12 h-12 bg-orange-100 rounded-lg flex items-center justify-center mb-4">
                  <i class="fas fa-expand-arrows-alt text-orange-600"></i>
                </div>
                <h4 class="font-semibold mb-3">理论推广挑战</h4>
                <p class="text-sm text-gray-600 mb-4">
                  如何将玩具任务的理论推广至大规模模型（如LLM）是一个巨大挑战
                </p>
                <div class="space-y-2">
                  <div class="bg-orange-50 p-2 rounded text-xs text-orange-700">
                    损失曲面极其复杂
                  </div>
                  <div class="bg-orange-50 p-2 rounded text-xs text-orange-700">
                    需要新的理论工具
                  </div>
                </div>
              </div>

              <div class="bg-white p-6 rounded-xl shadow-lg">
                <div class="w-12 h-12 bg-blue-100 rounded-lg flex items-center justify-center mb-4">
                  <i class="fas fa-cogs text-blue-600"></i>
                </div>
                <h4 class="font-semibold mb-3">可控偏置机制</h4>
                <p class="text-sm text-gray-600 mb-4">
                  设计可控且稳定的偏置引入机制，以诱导期望的泛化行为
                </p>
                <div class="space-y-2">
                  <div class="bg-blue-50 p-2 rounded text-xs text-blue-700">
                    自适应正则化
                  </div>
                  <div class="bg-blue-50 p-2 rounded text-xs text-blue-700">
                    基于元学习的方法
                  </div>
                </div>
              </div>

              <div class="bg-white p-6 rounded-xl shadow-lg">
                <div class="w-12 h-12 bg-green-100 rounded-lg flex items-center justify-center mb-4">
                  <i class="fas fa-eye text-green-600"></i>
                </div>
                <h4 class="font-semibold mb-3">多模态学习机遇</h4>
                <p class="text-sm text-gray-600 mb-4">
                  探索归纳偏置在多模态学习中的作用，设计促进跨模态泛化的偏置
                </p>
                <div class="space-y-2">
                  <div class="bg-green-50 p-2 rounded text-xs text-green-700">
                    模态无关表示
                  </div>
                  <div class="bg-green-50 p-2 rounded text-xs text-green-700">
                    语义空间对齐
                  </div>
                </div>
              </div>
            </div>

            <div class="mt-8 bg-gradient-to-r from-blue-50 to-purple-50 p-8 rounded-xl">
              <h4 class="font-semibold mb-4 text-center text-xl">研究方向展望</h4>
              <div class="grid md:grid-cols-2 gap-8">
                <div>
                  <h5 class="font-medium mb-3 text-blue-800">技术发展方向</h5>
                  <ul class="space-y-2 text-sm text-gray-700">
                    <li>• 开发自适应正则化方法</li>
                    <li>• 设计更智能的优化器</li>
                    <li>• 构建可控偏置引入机制</li>
                    <li>• 加速Grokking的训练策略</li>
                  </ul>
                </div>
                <div>
                  <h5 class="font-medium mb-3 text-purple-800">应用拓展方向</h5>
                  <ul class="space-y-2 text-sm text-gray-700">
                    <li>• 多模态学习中的偏置设计</li>
                    <li>• 科学计算领域的特定偏置</li>
                    <li>• AI for Science的结构编码</li>
                    <li>• 可控AI系统的偏置引导</li>
                  </ul>
                </div>
              </div>
            </div>
          </div>
        </div>
      </section>

      <!-- Conclusion -->
      <section class="py-16 bg-slate-50">
        <div class="container mx-auto px-8 max-w-4xl">
          <div class="text-center mb-12">
            <h2 class="text-3xl font-bold mb-6 serif">结论与启示</h2>
            <div class="w-24 h-1 bg-gradient-to-r from-blue-500 to-purple-500 mx-auto rounded"></div>
          </div>

          <div class="highlight-box">
            <h3 class="text-xl font-semibold mb-4 serif">核心总结</h3>
            <p class="text-lg leading-relaxed mb-4">
              归纳偏置为理解Grokking现象提供了强有力的理论框架。通过优化动态的二分性、电路竞争机制、Slingshot效应和复杂度偏好等多个角度，我们能够系统地解释模型从记忆到泛化的戏剧性转变。
            </p>
            <p class="text-gray-700">
              这一理解不仅揭示了深度学习泛化的内在机制，也为设计更智能、更可控的AI系统提供了宝贵的指导。未来的研究需要在理论和实践两个层面继续深入，以构建更完整的图景并将这些见解转化为可落地的技术。
            </p>
          </div>

          <div class="grid md:grid-cols-3 gap-6 mt-12">
            <div class="text-center p-6 bg-white rounded-xl shadow-lg">
              <div class="w-16 h-16 bg-blue-100 rounded-full flex items-center justify-center mx-auto mb-4">
                <i class="fas fa-lightbulb text-blue-600 text-2xl"></i>
              </div>
              <h4 class="font-semibold mb-2">理论价值</h4>
              <p class="text-sm text-gray-600">为理解深度学习泛化提供了新的理论视角</p>
            </div>

            <div class="text-center p-6 bg-white rounded-xl shadow-lg">
              <div class="w-16 h-16 bg-green-100 rounded-full flex items-center justify-center mx-auto mb-4">
                <i class="fas fa-cogs text-green-600 text-2xl"></i>
              </div>
              <h4 class="font-semibold mb-2">实践指导</h4>
              <p class="text-sm text-gray-600">为模型训练和优化提供了实用的策略指导</p>
            </div>

            <div class="text-center p-6 bg-white rounded-xl shadow-lg">
              <div class="w-16 h-16 bg-purple-100 rounded-full flex items-center justify-center mx-auto mb-4">
                <i class="fas fa-rocket text-purple-600 text-2xl"></i>
              </div>
              <h4 class="font-semibold mb-2">未来潜力</h4>
              <p class="text-sm text-gray-600">为构建更强大的AI系统指明了发展方向</p>
            </div>
          </div>
        </div>
      </section>
    </main>

    <script>
        // Initialize Mermaid
        mermaid.initialize({
            startOnLoad: true,
            theme: 'base',
            themeVariables: {
                primaryColor: '#2563eb',
                primaryTextColor: '#1e293b',
                primaryBorderColor: '#64748b',
                lineColor: '#64748b',
                secondaryColor: '#f8fafc',
                tertiaryColor: '#e2e8f0',
                background: '#ffffff',
                mainBkg: '#ffffff',
                secondBkg: '#f8fafc',
                tertiaryBkg: '#e2e8f0',
                nodeBorder: '#64748b',
                clusterBkg: '#f8fafc',
                edgeLabelBackground: '#ffffff',
                nodeTextColor: '#1e293b'
            },
            flowchart: {
                useMaxWidth: false,
                htmlLabels: true,
                curve: 'basis',
                padding: 20
            },
            sequence: {
                useMaxWidth: false,
                wrap: true
            },
            gantt: {
                useMaxWidth: false
            }
        });

        // Initialize Mermaid Controls for zoom and pan
        function initializeMermaidControls() {
            const containers = document.querySelectorAll('.mermaid-container');

            containers.forEach(container => {
            const mermaidElement = container.querySelector('.mermaid');
            let scale = 1;
            let isDragging = false;
            let startX, startY, translateX = 0, translateY = 0;

            // 触摸相关状态
            let isTouch = false;
            let touchStartTime = 0;
            let initialDistance = 0;
            let initialScale = 1;
            let isPinching = false;

            // Zoom controls
            const zoomInBtn = container.querySelector('.zoom-in');
            const zoomOutBtn = container.querySelector('.zoom-out');
            const resetBtn = container.querySelector('.reset-zoom');
            const fullscreenBtn = container.querySelector('.fullscreen');

            function updateTransform() {
                mermaidElement.style.transform = `translate(${translateX}px, ${translateY}px) scale(${scale})`;

                if (scale > 1) {
                container.classList.add('zoomed');
                } else {
                container.classList.remove('zoomed');
                }

                mermaidElement.style.cursor = isDragging ? 'grabbing' : 'grab';
            }

            if (zoomInBtn) {
                zoomInBtn.addEventListener('click', () => {
                scale = Math.min(scale * 1.25, 4);
                updateTransform();
                });
            }

            if (zoomOutBtn) {
                zoomOutBtn.addEventListener('click', () => {
                scale = Math.max(scale / 1.25, 0.3);
                if (scale <= 1) {
                    translateX = 0;
                    translateY = 0;
                }
                updateTransform();
                });
            }

            if (resetBtn) {
                resetBtn.addEventListener('click', () => {
                scale = 1;
                translateX = 0;
                translateY = 0;
                updateTransform();
                });
            }

            if (fullscreenBtn) {
                fullscreenBtn.addEventListener('click', () => {
                if (container.requestFullscreen) {
                    container.requestFullscreen();
                } else if (container.webkitRequestFullscreen) {
                    container.webkitRequestFullscreen();
                } else if (container.msRequestFullscreen) {
                    container.msRequestFullscreen();
                }
                });
            }

            // Mouse Events
            mermaidElement.addEventListener('mousedown', (e) => {
                if (isTouch) return; // 如果是触摸设备，忽略鼠标事件

                isDragging = true;
                startX = e.clientX - translateX;
                startY = e.clientY - translateY;
                mermaidElement.style.cursor = 'grabbing';
                updateTransform();
                e.preventDefault();
            });

            document.addEventListener('mousemove', (e) => {
                if (isDragging && !isTouch) {
                translateX = e.clientX - startX;
                translateY = e.clientY - startY;
                updateTransform();
                }
            });

            document.addEventListener('mouseup', () => {
                if (isDragging && !isTouch) {
                isDragging = false;
                mermaidElement.style.cursor = 'grab';
                updateTransform();
                }
            });

            document.addEventListener('mouseleave', () => {
                if (isDragging && !isTouch) {
                isDragging = false;
                mermaidElement.style.cursor = 'grab';
                updateTransform();
                }
            });

            // 获取两点之间的距离
            function getTouchDistance(touch1, touch2) {
                return Math.hypot(
                touch2.clientX - touch1.clientX,
                touch2.clientY - touch1.clientY
                );
            }

            // Touch Events - 触摸事件处理
            mermaidElement.addEventListener('touchstart', (e) => {
                isTouch = true;
                touchStartTime = Date.now();

                if (e.touches.length === 1) {
                // 单指拖动
                isPinching = false;
                isDragging = true;

                const touch = e.touches[0];
                startX = touch.clientX - translateX;
                startY = touch.clientY - translateY;

                } else if (e.touches.length === 2) {
                // 双指缩放
                isPinching = true;
                isDragging = false;

                const touch1 = e.touches[0];
                const touch2 = e.touches[1];
                initialDistance = getTouchDistance(touch1, touch2);
                initialScale = scale;
                }

                e.preventDefault();
            }, { passive: false });

            mermaidElement.addEventListener('touchmove', (e) => {
                if (e.touches.length === 1 && isDragging && !isPinching) {
                // 单指拖动
                const touch = e.touches[0];
                translateX = touch.clientX - startX;
                translateY = touch.clientY - startY;
                updateTransform();

                } else if (e.touches.length === 2 && isPinching) {
                // 双指缩放
                const touch1 = e.touches[0];
                const touch2 = e.touches[1];
                const currentDistance = getTouchDistance(touch1, touch2);

                if (initialDistance > 0) {
                    const newScale = Math.min(Math.max(
                    initialScale * (currentDistance / initialDistance),
                    0.3
                    ), 4);
                    scale = newScale;
                    updateTransform();
                }
                }

                e.preventDefault();
            }, { passive: false });

            mermaidElement.addEventListener('touchend', (e) => {
                // 重置状态
                if (e.touches.length === 0) {
                isDragging = false;
                isPinching = false;
                initialDistance = 0;

                // 延迟重置isTouch，避免鼠标事件立即触发
                setTimeout(() => {
                    isTouch = false;
                }, 100);
                } else if (e.touches.length === 1 && isPinching) {
                // 从双指变为单指，切换为拖动模式
                isPinching = false;
                isDragging = true;

                const touch = e.touches[0];
                startX = touch.clientX - translateX;
                startY = touch.clientY - translateY;
                }

                updateTransform();
            });

            mermaidElement.addEventListener('touchcancel', (e) => {
                isDragging = false;
                isPinching = false;
                initialDistance = 0;

                setTimeout(() => {
                isTouch = false;
                }, 100);

                updateTransform();
            });

            // Enhanced wheel zoom with better center point handling
            container.addEventListener('wheel', (e) => {
                e.preventDefault();
                const rect = container.getBoundingClientRect();
                const centerX = rect.width / 2;
                const centerY = rect.height / 2;

                const delta = e.deltaY > 0 ? 0.9 : 1.1;
                const newScale = Math.min(Math.max(scale * delta, 0.3), 4);

                // Adjust translation to zoom towards center
                if (newScale !== scale) {
                const scaleDiff = newScale / scale;
                translateX = translateX * scaleDiff;
                translateY = translateY * scaleDiff;
                scale = newScale;

                if (scale <= 1) {
                    translateX = 0;
                    translateY = 0;
                }

                updateTransform();
                }
            });

            // Initialize display
            updateTransform();
            });
        }

        // Smooth scrolling for anchor links
        document.querySelectorAll('a[href^="#"]').forEach(anchor => {
            anchor.addEventListener('click', function (e) {
                e.preventDefault();
                const target = document.querySelector(this.getAttribute('href'));
                if (target) {
                    target.scrollIntoView({
                        behavior: 'smooth',
                        block: 'start'
                    });
                }
            });
        });

        // Active section highlighting in TOC
        const sections = document.querySelectorAll('section[id]');
        const tocLinks = document.querySelectorAll('.toc-link');

        function updateActiveSection() {
            let current = '';
            sections.forEach(section => {
                const sectionTop = section.offsetTop;
                const sectionHeight = section.clientHeight;
                if (pageYOffset >= sectionTop - 200) {
                    current = section.getAttribute('id');
                }
            });

            tocLinks.forEach(link => {
                link.classList.remove('active');
                if (link.getAttribute('href') === `#${current}`) {
                    link.classList.add('active');
                }
            });
        }

        window.addEventListener('scroll', updateActiveSection);
        updateActiveSection();

        // Mobile TOC toggle
        const tocToggle = document.createElement('button');
        tocToggle.innerHTML = '<i class="fas fa-bars"></i>';
        tocToggle.className = 'fixed top-4 left-4 z-50 lg:hidden bg-white p-3 rounded-full shadow-lg';
        tocToggle.onclick = () => {
            document.querySelector('.toc-sidebar').classList.toggle('open');
        };
        document.body.appendChild(tocToggle);

        // Close TOC when clicking outside on mobile
        document.addEventListener('click', (e) => {
            const sidebar = document.querySelector('.toc-sidebar');
            if (window.innerWidth < 1024 && sidebar.classList.contains('open') && 
                !sidebar.contains(e.target) && e.target !== tocToggle) {
                sidebar.classList.remove('open');
            }
        });

        // Initialize mermaid controls
        initializeMermaidControls();
    </script>
  

</body></html>                    

讨论回复

1 条回复

✨步子哥 (steper) #1

12-22 08:22

                                        ![3617962-20250315000618360-853825764.png](https://s2.loli.net/2025/12/22/IZbMWXfG29cQVTm.png)                                    

需要登录才能发表回复

登录注册

LLM的顿悟现象

讨论回复

推荐