ai-agent-book 精选快照（<2MB 代码与文档，来自 github.com/bojieli/ai-agent-book）

chapter5/paper-to-ppt/validation/runs/exp5-4-real-pdf-both-20260730-v5/dual_round1_review.json

@@ -0,0 +1,54 @@
+{
+  "overall_score": 65,
+  "pass": false,
+  "issues": [
+    {
+      "page": 2,
+      "issue_type": "overcrowded",
+      "severity": "medium",
+      "suggestion": "要点数量达6条，超出建议的5条上限，建议将最后1-2条拆分至下一页或合并相似内容"
+    },
+    {
+      "page": 3,
+      "issue_type": "overcrowded",
+      "severity": "medium",
+      "suggestion": "要点数量达7条（含子要点），内容拥挤，建议将Recurrent models和Convolutional models分拆为两页展示"
+    },
+    {
+      "page": 13,
+      "issue_type": "readability",
+      "severity": "high",
+      "suggestion": "Learning rate公式存在LaTeX渲染错误（显示为“d_{text{model}}^{-0.5} \\cdot \\min(\\text{step_num}^{-0.5}, \\text{step_num} \\cdot \\text{warmup_steps}^{-1.5})”），需修正公式格式确保正确显示"
+    },
+    {
+      "page": 15,
+      "issue_type": "layout",
+      "severity": "low",
+      "suggestion": "底部要点符号不一致（混合使用“-”和“■”），建议统一使用项目符号“■”保持格式统一"
+    },
+    {
+      "page": 17,
+      "issue_type": "image_size",
+      "severity": "medium",
+      "suggestion": "注意力可视化图尺寸过小，文字和连线细节模糊不清，建议放大图片至页面70%以上面积或分拆为多页展示"
+    },
+    {
+      "page": 18,
+      "issue_type": "image_size",
+      "severity": "medium",
+      "suggestion": "指代消解可视化图尺寸不足，紫色线条和文字难以辨认，建议单独放大展示或增加图例说明"
+    },
+    {
+      "page": 19,
+      "issue_type": "overcrowded",
+      "severity": "medium",
+      "suggestion": "Limitations（3条）和Future work（4条）共7个要点，内容过于密集，建议分拆为“局限性”和“未来工作”两页分别展示"
+    },
+    {
+      "page": 20,
+      "issue_type": "overcrowded",
+      "severity": "medium",
+      "suggestion": "结论要点达6条，超出建议数量，建议删除“Code available...”等次要信息或拆分至单独的“资源链接”页"
+    }
+  ]
+}

chapter5/paper-to-ppt/validation/runs/exp5-4-real-pdf-both-20260730-v5/dual_round1_slides.md

@@ -0,0 +1,251 @@
+---
+theme: default
+---
+
+# Attention Is All You Need
+## A Revolutionary Architecture for Sequence Transduction
+
+Ashish Vaswani et al.  
+NIPS 2017
+
+---
+
+## Abstract
+
+- Proposes Transformer: first model based solely on attention mechanisms
+- Dispenses with recurrence and convolutions entirely
+- Achieves superior quality while being more parallelizable
+- 28.4 BLEU on WMT 2014 English-to-German (↑2+ BLEU)
+- 41.8 BLEU on WMT 2014 English-to-French (new state-of-the-art)
+- Generalizes well to other tasks like English constituency parsing
+
+---
+
+## Background: The Problem with Traditional Approaches
+
+- **Recurrent models (RNN/LSTM/GRU)**
+  - Inherently sequential computation
+  - Limited parallelization
+  - Difficult to learn long-range dependencies
+
+- **Convolutional models**
+  - Fixed kernel size limits long-range dependencies
+  - O(n/k) or O(logk(n)) operations for distant connections
+  - Less efficient than attention for sequence tasks
+
+---
+
+## Key Innovation: Attention as the Core Mechanism
+
+- **Self-attention** connects all positions with constant operations
+- **Parallelization** possible across sequence positions
+- **Long-range dependencies** modeled directly
+- **Interpretability** through attention distributions
+- **Computational efficiency** for typical sequence lengths
+
+---
+
+## Transformer Model Architecture
+
+<img src="/paper_figure_1_transformer.png" class="h-[560px] w-full object-contain" />
+
+The overall encoder-decoder structure with stacked self-attention and feed-forward layers.
+
+---
+
+## Encoder Architecture
+
+- Stack of 6 identical layers
+- Each layer has two sub-layers:
+  1. Multi-head self-attention mechanism
+  2. Position-wise fully connected feed-forward network
+- Residual connections around each sub-layer
+- Layer normalization after each sub-layer
+- All sub-layers produce outputs of dimension dmodel = 512
+
+---
+
+## Decoder Architecture
+
+- Stack of 6 identical layers
+- Three sub-layers per layer:
+  1. Masked multi-head self-attention
+  2. Multi-head attention over encoder outputs
+  3. Position-wise fully connected feed-forward network
+- Residual connections and layer normalization
+- Masking prevents attending to future positions
+- Output embeddings offset by one position
+
+---
+
+## Attention Mechanism
+
+- Maps query and key-value pairs to output
+- Output = weighted sum of values
+- Weights computed by compatibility function of query and keys
+
+**Scaled Dot-Product Attention:**
+$$\text{Attention}(Q, K, V) = \text{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right)V$$
+
+- Scaling prevents gradients from becoming too small
+- More efficient than additive attention in practice
+
+---
+
+## Multi-Head Attention
+
+- Projects queries, keys, values h times with different linear projections
+- Performs attention in parallel on projected subspaces
+- Concatenates results and projects again
+
+$$\text{MultiHead}(Q, K, V) = \text{Concat}(\text{head}_1, ..., \text{head}_h)W^O$$
+where $\text{head}_i = \text{Attention}(QW_i^Q, KW_i^K, VW_i^V)$
+
+- h = 8 parallel attention heads in base model
+- $d_k = d_v = d_{\text{model}}/h = 64$
+
+---
+
+## Three Applications of Attention
+
+1. **Encoder-decoder attention**
+   - Queries from decoder, keys/values from encoder
+   - Allows decoder to attend to all input positions
+
+2. **Encoder self-attention**
+   - All keys, values, queries from previous encoder layer
+   - Each position attends to all positions in previous layer
+
+3. **Decoder self-attention**
+   - Each position attends to previous positions in decoder
+   - Masking prevents attending to future positions
+
+---
+
+## Positional Encoding
+
+- Injects sequence order information since no recurrence/convolution
+- Added to input embeddings (same dimension dmodel)
+- Uses sine and cosine functions of different frequencies:
+
+$$PE_{(pos,2i)} = \sin(pos/10000^{2i/d_{\text{model}}})$$
+$$PE_{(pos,2i+1)} = \cos(pos/10000^{2i/d_{\text{model}}})$$
+
+- Allows model to learn relative position relationships
+- Performed as well as learned positional embeddings
+
+---
+
+## Why Self-Attention?
+
+| Aspect | Self-Attention | Recurrent | Convolutional |
+|--------|----------------|-----------|---------------|
+| Complexity | O(n²·d) | O(n·d²) | O(k·n·d²) |
+| Parallelization | O(1) | O(n) | O(1) |
+| Long-range paths | O(1) | O(n) | O(logk(n)) |
+
+- More parallelizable than RNNs
+- Better at capturing long-range dependencies than CNNs
+- More efficient for typical sequence lengths (n < d)
+- More interpretable through attention weights
+
+---
+
+## Training Details
+
+- **Datasets**: WMT 2014 EN-DE (4.5M) and EN-FR (36M)
+- **Batching**: ~25,000 source and target tokens per batch
+- **Hardware**: 8 NVIDIA P100 GPUs
+- **Optimizer**: Adam (β1=0.9, β2=0.98, ϵ=10⁻⁹)
+- **Learning rate**: $d_{\text{model}}^{-0.5} \cdot \min(\text{step_num}^{-0.5}, \text{step_num} \cdot \text{warmup_steps}^{-1.5})$
+- **Regularization**: Residual dropout (Pdrop=0.1), label smoothing (ϵls=0.1)
+
+---
+
+## Machine Translation Results
+
+| Model | EN-DE BLEU | EN-FR BLEU | Training Cost (FLOPs) |
+|-------|------------|------------|-----------------------|
+| GNMT + RL Ensemble | 26.30 | 41.16 | 1.8×10²⁰ / 1.1×10²¹ |
+| ConvS2S Ensemble | 26.36 | 41.29 | 7.7×10¹⁹ / 1.2×10²¹ |
+| **Transformer (big)** | **28.4** | **41.8** | **2.3×10¹⁹** |
+
+- Transformer (big) outperforms all previous models
+- Achieves new state-of-the-art with significantly lower training cost
+- Trained in 3.5 days on 8 GPUs (vs. weeks for competitors)
+
+---
+
+## Model Architecture Ablations
+
+| Variation | Dev PPL | Dev BLEU |
+|-----------|---------|----------|
+| Base model | 4.92 | 25.8 |
+| Single attention head | 5.29 | 24.9 |
+| 32 attention heads | 5.01 | 25.4 |
+| No dropout | 4.67 | 25.3 |
+| Learned positional embeddings | 4.92 | 25.7 |
+| Big model | 4.33 | 26.4 |
+
+-多头注意力优于单头注意力
+- dropout对防止过拟合至关重要
+- 正弦位置编码与学习的位置编码效果相当
+- 增大模型尺寸(dmodel=1024, dff=4096)提升性能
+
+---
+
+## Generalization to English Constituency Parsing
+
+| Parser | Training | WSJ 23 F1 |
+|--------|----------|-----------|
+| Petrov et al. (2006) | WSJ only | 90.4 |
+| Dyer et al. (2016) | WSJ only | 91.7 |
+| **Transformer (4 layers)** | **WSJ only** | **91.3** |
+| Vinyals & Kaiser (2014) | Semi-supervised | 92.1 |
+| **Transformer (4 layers)** | **Semi-supervised** | **92.7** |
+
+- Transformer performs well without task-specific modifications
+- Outperforms RNN approaches in small-data regimes
+- Achieves 92.7 F1 in semi-supervised setting
+
+---
+
+## Attention Visualization: Long-Distance Dependencies
+
+<img src="/paper_figure_3_long_distance.png" class="h-[560px] w-full object-contain" />
+
+Encoder self-attention in layer 5 showing attention to distant dependency of the verb "making".
+
+---
+
+## Attention Visualization: Anaphora Resolution
+
+<img src="/paper_figure_4_anaphora.png" class="h-[560px] w-full object-contain" />
+
+Attention heads involved in resolving the pronoun "its" to its referent "The Law".
+
+---
+
+## Limitations and Future Work
+
+- **Limitations**:
+  - Quadratic complexity in sequence length
+  - Less effective for very long sequences
+  - Still generates output sequentially
+
+- **Future work**:
+  - Local, restricted attention mechanisms
+  - Applications to other modalities (images, audio, video)
+  - Less sequential generation approaches
+  - Efficient handling of large inputs/outputs
+
+---
+
+## Conclusion
+
+- Transformer achieves state-of-the-art results in machine translation
+- Eliminates recurrence and convolution in favor of self-attention
+- Significantly faster training through parallelization
+- Generalizes well to other sequence tasks like constituency parsing
+- Paves the way for modern attention-based models in NLP
+- Code available at https://github.com/tensorflow/tensor2tensor

chapter5/paper-to-ppt/validation/runs/exp5-4-real-pdf-both-20260730-v5/dual_round2_review.json

@@ -0,0 +1,48 @@
+{
+  "overall_score": 75,
+  "pass": false,
+  "issues": [
+    {
+      "page": 3,
+      "issue_type": "overcrowded",
+      "severity": "medium",
+      "suggestion": "将Recurrent Models和Convolutional Models分为两页展示，或每个部分保留2个最关键要点"
+    },
+    {
+      "page": 6,
+      "issue_type": "overcrowded",
+      "severity": "high",
+      "suggestion": "将Encoder和Decoder架构分为两页，每部分控制在4个要点以内，删除重复的\"Residual connections + layer normalization\""
+    },
+    {
+      "page": 10,
+      "issue_type": "overcrowded",
+      "severity": "medium",
+      "suggestion": "将Training Setup拆分为数据/硬件和优化策略两页，或删除学习率公式的数学表达仅保留文字描述"
+    },
+    {
+      "page": 12,
+      "issue_type": "readability",
+      "severity": "low",
+      "suggestion": "将底部总结文字拆分为两个独立项目符号，使用更清晰的分隔符替代当前的黑色方块符号"
+    },
+    {
+      "page": 14,
+      "issue_type": "image_size",
+      "severity": "medium",
+      "suggestion": "放大注意力可视化图表，确保轴标签文字清晰可辨，或裁剪部分padding区域聚焦核心可视化内容"
+    },
+    {
+      "page": 15,
+      "issue_type": "image_size",
+      "severity": "medium",
+      "suggestion": "放大指代消解可视化图表，增加线条对比度，或仅保留最具代表性的1-2个注意力头可视化"
+    },
+    {
+      "page": 16,
+      "issue_type": "overcrowded",
+      "severity": "medium",
+      "suggestion": "将Limitations和Future Work分为两页，或每个部分精简为2-3个核心要点"
+    }
+  ]
+}

chapter5/paper-to-ppt/validation/runs/exp5-4-real-pdf-both-20260730-v5/dual_round2_slides.md

@@ -0,0 +1,202 @@
+---
+theme: default
+---
+
+# Attention Is All You Need
+## A Revolutionary Architecture for Sequence Transduction
+
+Ashish Vaswani et al.  
+NIPS 2017
+
+---
+
+## Abstract
+
+- First sequence transduction model based solely on attention
+- Dispenses with recurrence and convolutions entirely
+- More parallelizable and faster to train than traditional models
+- New state-of-the-art: 28.4 BLEU (EN-DE) and 41.8 BLEU (EN-FR)
+- Generalizes well to English constituency parsing
+
+---
+
+## Background: Limitations of Traditional Models
+
+### Recurrent Models (RNN/LSTM/GRU)
+- Inherently sequential computation
+- Limited parallelization capability
+- O(n) sequential operations for long sequences
+
+### Convolutional Models
+- Fixed kernel size restricts context
+- O(k·n·d²) complexity with kernel size k
+- Logarithmic path length for distant dependencies
+
+---
+
+## Key Innovation: Self-Attention Mechanism
+
+- Connects all positions with constant operations
+- Enables parallel computation across sequence
+- Directly models long-range dependencies
+- More efficient than RNN/CNN for typical sequence lengths
+- Provides interpretable attention distributions
+
+---
+
+## Transformer Model Architecture
+
+<img src="/paper_figure_1_transformer.png" class="h-[560px] w-full object-contain" />
+
+Encoder-decoder structure with stacked self-attention and feed-forward layers.
+
+---
+
+## Encoder & Decoder Architecture
+
+### Encoder (6 identical layers)
+- **Sub-layer 1**: Multi-head self-attention
+- **Sub-layer 2**: Position-wise feed-forward network
+- Residual connections + layer normalization
+- Output dimension: dmodel = 512
+
+### Decoder (6 identical layers)
+- **Sub-layer 1**: Masked multi-head self-attention
+- **Sub-layer 2**: Encoder-decoder attention
+- **Sub-layer 3**: Position-wise feed-forward network
+- Residual connections + layer normalization
+- Masking prevents future position attention
+
+---
+
+## Attention Mechanisms
+
+### Scaled Dot-Product Attention
+$$\text{Attention}(Q, K, V) = \text{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right)V$$
+- Scaling avoids gradient vanishing for large dk
+- More efficient than additive attention
+
+### Multi-Head Attention
+$$\text{MultiHead}(Q, K, V) = \text{Concat}(\text{head}_1,...,\text{head}_h)W^O$$
+- h=8 parallel heads, dk=dv=64
+- Captures diverse dependency patterns
+
+---
+
+## Positional Encoding
+
+- Injects sequence order information (no recurrence/convolution)
+- Added to input embeddings (same dmodel dimension)
+- Uses sine/cosine functions with varying frequencies:
+  $$PE_{(pos,2i)} = \sin(pos/10000^{2i/d_{\text{model}}})$$
+  $$PE_{(pos,2i+1)} = \cos(pos/10000^{2i/d_{\text{model}}})$$
+- Enables learning of relative position relationships
+
+---
+
+## Why Self-Attention?
+
+| Aspect               | Self-Attention | Recurrent | Convolutional |
+|----------------------|----------------|-----------|---------------|
+| Complexity           | O(n²·d)        | O(n·d²)   | O(k·n·d²)     |
+| Parallelization      | O(1)           | O(n)      | O(1)          |
+| Long-range path length | O(1)         | O(n)      | O(logk(n))    |
+
+- Superior parallelization and dependency modeling
+
+---
+
+## Training Setup
+
+### Data & Hardware
+- WMT 2014 EN-DE (4.5M) and EN-FR (36M)
+- 8 NVIDIA P100 GPUs, 0.4s/step (base), 1.0s/step (big)
+
+### Optimization
+- Adam (β1=0.9, β2=0.98, ϵ=10⁻⁹)
+- Learning rate: $d_{\text{model}}^{-0.5} \cdot \min(\text{step\_num}^{-0.5}, \text{step\_num} \cdot \text{warmup\_steps}^{-1.5})$
+
+### Regularization
+- Residual dropout (Pdrop=0.1)
+- Label smoothing (ϵls=0.1)
+
+---
+
+## Machine Translation Results
+
+| Model | EN-DE BLEU | EN-FR BLEU | Training Cost (FLOPs) |
+|-------|------------|------------|-----------------------|
+| GNMT + RL Ensemble | 26.30 | 41.16 | 1.8×10²⁰ / 1.1×10²¹ |
+| ConvS2S Ensemble | 26.36 | 41.29 | 7.7×10¹⁹ / 1.2×10²¹ |
+| **Transformer (big)** | **28.4** | **41.8** | **2.3×10¹⁹** |
+
+- New state-of-the-art with 4× lower training cost
+
+---
+
+## Model Ablations (EN-DE Dev Set)
+
+| Variation | Dev PPL | Dev BLEU |
+|-----------|---------|----------|
+| Base model | 4.92 | 25.8 |
+| Single attention head | 5.29 | 24.9 |
+| No dropout | 4.67 | 25.3 |
+| Learned positional embeddings | 4.92 | 25.7 |
+| Big model | 4.33 | 26.4 |
+
+■ Multi-head attention and dropout critical for performance
+■ Sinusoidal positional encoding ≈ learned embeddings
+
+---
+
+## Generalization to Constituency Parsing
+
+| Parser | Training | WSJ 23 F1 |
+|--------|----------|-----------|
+| Dyer et al. (2016) | WSJ only | 91.7 |
+| **Transformer (4 layers)** | **WSJ only** | **91.3** |
+| Vinyals & Kaiser (2014) | Semi-supervised | 92.1 |
+| **Transformer (4 layers)** | **Semi-supervised** | **92.7** |
+
+- Strong performance without task-specific modifications
+
+---
+
+## Attention Visualization: Long-Distance Dependencies
+
+<img src="/paper_figure_3_long_distance.png" class="h-[580px] w-full object-contain" />
+
+Encoder self-attention (layer 5) tracking "making...more difficult" dependency.
+
+---
+
+## Attention Visualization: Anaphora Resolution
+
+<img src="/paper_figure_4_anaphora.png" class="h-[580px] w-full object-contain" />
+
+Attention heads resolving "its" to referent "The Law".
+
+---
+
+## Limitations & Future Work
+
+### Limitations
+- Quadratic complexity in sequence length
+- Less efficient for very long sequences
+- Still sequential in generation
+
+### Future Work
+- Local/restricted attention mechanisms
+- Extension to other modalities (images, audio)
+- Non-sequential generation approaches
+- Efficient handling of large inputs/outputs
+
+---
+
+## Conclusion
+
+- Transformer replaces recurrence/convolution with self-attention
+- Sets new state-of-the-art in machine translation
+- Faster training via parallelization
+- Generalizes well to diverse sequence tasks
+- Foundation for modern attention-based NLP models

chapter5/paper-to-ppt/validation/runs/exp5-4-real-pdf-both-20260730-v5/dual_round3_review.json

@@ -0,0 +1,24 @@
+{
+  "overall_score": 85,
+  "pass": false,
+  "issues": [
+    {
+      "page": 14,
+      "issue_type": "readability",
+      "severity": "medium",
+      "suggestion": "可视化图中的文字字号过小，建议放大文字或调整图表布局以提高可读性"
+    },
+    {
+      "page": 15,
+      "issue_type": "readability",
+      "severity": "medium",
+      "suggestion": "可视化图中的文字字号过小，建议放大文字或调整图表布局以提高可读性"
+    },
+    {
+      "page": 16,
+      "issue_type": "overcrowded",
+      "severity": "low",
+      "suggestion": "当前页面包含6个要点，略超出建议的5个要点限制，可考虑将Limitations和Future Work分为两页展示"
+    }
+  ]
+}

chapter5/paper-to-ppt/validation/runs/exp5-4-real-pdf-both-20260730-v5/dual_round3_slides.md

@@ -0,0 +1,184 @@
+---
+theme: default
+---
+
+# Attention Is All You Need
+## A Revolutionary Architecture for Sequence Transduction
+
+Ashish Vaswani et al.  
+NIPS 2017
+
+---
+
+## Abstract
+
+- First sequence transduction model based solely on attention
+- Dispenses with recurrence and convolutions entirely
+- More parallelizable and faster to train than traditional models
+- New state-of-the-art: 28.4 BLEU (EN-DE) and 41.8 BLEU (EN-FR)
+- Generalizes well to English constituency parsing
+
+---
+
+## Background: Limitations of Traditional Models
+
+### Recurrent Models (RNN/LSTM/GRU)
+- Inherently sequential computation → limited parallelization
+- O(n) sequential operations for long-range dependencies
+
+### Convolutional Models
+- Fixed kernel size restricts context → requires multiple layers
+- Logarithmic path length for distant connections
+
+---
+
+## Key Innovation: Self-Attention Mechanism
+
+- Connects all positions with constant operations
+- Enables parallel computation across sequence
+- Directly models long-range dependencies
+- More efficient than RNN/CNN for typical sequence lengths
+
+---
+
+## Transformer Model Architecture
+
+<img src="/paper_figure_1_transformer.png" class="h-[560px] w-full object-contain" />
+
+Encoder-decoder structure with stacked self-attention and feed-forward layers.
+
+---
+
+## Encoder Architecture
+
+- Stack of 6 identical layers
+- Each layer has two sub-layers:
+  1. Multi-head self-attention mechanism
+  2. Position-wise feed-forward network
+- Residual connections + layer normalization
+- Output dimension: dmodel = 512
+
+---
+
+## Decoder Architecture
+
+- Stack of 6 identical layers
+- Three sub-layers per layer:
+  1. Masked multi-head self-attention (prevents future positions)
+  2. Encoder-decoder attention (queries from decoder, keys/values from encoder)
+  3. Position-wise feed-forward network
+- Residual connections + layer normalization
+
+---
+
+## Attention Mechanisms
+
+### Scaled Dot-Product Attention
+$$\text{Attention}(Q, K, V) = \text{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right)V$$
+- Scaling avoids gradient vanishing for large dk
+
+### Multi-Head Attention
+$$\text{MultiHead}(Q, K, V) = \text{Concat}(\text{head}_1,...,\text{head}_h)W^O$$
+- h=8 parallel heads, dk=dv=64 → captures diverse patterns
+
+---
+
+## Positional Encoding
+
+- Injects sequence order information (no recurrence/convolution)
+- Added to input embeddings (same dmodel dimension)
+- Uses sine/cosine functions with varying frequencies:
+  $$PE_{(pos,2i)} = \sin(pos/10000^{2i/d_{\text{model}}})$$
+  $$PE_{(pos,2i+1)} = \cos(pos/10000^{2i/d_{\text{model}}})$$
+- Enables learning of relative position relationships
+
+---
+
+## Why Self-Attention?
+
+| Aspect               | Self-Attention | Recurrent | Convolutional |
+|----------------------|----------------|-----------|---------------|
+| Complexity           | O(n²·d)        | O(n·d²)   | O(k·n·d²)     |
+| Parallelization      | O(1)           | O(n)      | O(1)          |
+| Long-range path length | O(1)         | O(n)      | O(logk(n))    |
+
+- Superior parallelization and dependency modeling
+
+---
+
+## Training Setup
+
+### Data & Hardware
+- WMT 2014 EN-DE (4.5M) and EN-FR (36M) sentence pairs
+- 8 NVIDIA P100 GPUs, 12h (base)/3.5d (big model) training
+
+### Optimization
+- Adam (β1=0.9, β2=0.98, ϵ=10⁻⁹) with linear warmup (4000 steps) + inverse square root decay
+- Regularization: residual dropout (0.1), label smoothing (0.1)
+
+---
+
+## Machine Translation Results
+
+| Model | EN-DE BLEU | EN-FR BLEU | Training Cost (FLOPs) |
+|-------|------------|------------|-----------------------|
+| GNMT + RL Ensemble | 26.30 | 41.16 | 1.8×10²⁰ / 1.1×10²¹ |
+| ConvS2S Ensemble | 26.36 | 41.29 | 7.7×10¹⁹ / 1.2×10²¹ |
+| **Transformer (big)** | **28.4** | **41.8** | **2.3×10¹⁹** |
+
+- New state-of-the-art with 4× lower training cost
+
+---
+
+## Model Ablations & Generalization
+
+### Key Ablations (EN-DE Dev Set)
+| Variation | Dev BLEU | Insight |
+|-----------|----------|---------|
+| Single head | 24.9 | Multi-head critical |
+| No dropout | 25.3 | Regularization needed |
+| Learned pos encoding | 25.7 | Sinusoidal ≈ learned |
+
+### Constituency Parsing
+- 91.3 F1 (WSJ only) vs. 91.7 (state-of-the-art)
+- 92.7 F1 (semi-supervised) → strong generalization
+
+---
+
+## Attention Visualization: Long-Distance Dependencies
+
+<img src="/paper_figure_3_long_distance.png" class="h-[600px] w-full object-contain" />
+
+Encoder self-attention (layer 5) tracking "making...more difficult" dependency.
+
+---
+
+## Attention Visualization: Anaphora Resolution
+
+<img src="/paper_figure_4_anaphora.png" class="h-[600px] w-full object-contain" />
+
+Attention heads resolving "its" to referent "The Law".
+
+---
+
+## Limitations & Future Work
+
+### Limitations
+- Quadratic complexity in sequence length
+- Less efficient for very long sequences
+- Still sequential in generation
+
+### Future Work
+- Local/restricted attention mechanisms
+- Extension to other modalities (images, audio)
+- Non-sequential generation approaches
+
+---
+
+## Conclusion
+
+- Transformer replaces recurrence/convolution with self-attention
+- Sets new state-of-the-art in machine translation
+- Faster training via parallelization
+- Generalizes well to diverse sequence tasks
+- Foundation for modern attention-based NLP models

chapter5/paper-to-ppt/validation/runs/exp5-4-real-pdf-both-20260730-v5/dual_round4_review.json

@@ -0,0 +1,18 @@
+{
+  "overall_score": 90,
+  "pass": false,
+  "issues": [
+    {
+      "page": 14,
+      "issue_type": "image_size",
+      "severity": "medium",
+      "suggestion": "放大注意力可视化图，确保文字标签清晰可辨，避免因图片过小影响观众理解长距离依赖关系的展示效果"
+    },
+    {
+      "page": 15,
+      "issue_type": "image_size",
+      "severity": "medium",
+      "suggestion": "增大指代消解注意力可视化图的尺寸，确保文本标签和连接线清晰可见，提升观众对注意力模式的识别能力"
+    }
+  ]
+}

chapter5/paper-to-ppt/validation/runs/exp5-4-real-pdf-both-20260730-v5/dual_round4_slides.md

@@ -0,0 +1,187 @@
+---
+theme: default
+---
+
+# Attention Is All You Need
+## A Revolutionary Architecture for Sequence Transduction
+
+Ashish Vaswani et al.  
+NIPS 2017
+
+---
+
+## Abstract
+
+- First sequence transduction model based solely on attention
+- Dispenses with recurrence and convolutions entirely
+- More parallelizable and faster to train than traditional models
+- New state-of-the-art: 28.4 BLEU (EN-DE) and 41.8 BLEU (EN-FR)
+- Generalizes well to English constituency parsing
+
+---
+
+## Background: Limitations of Traditional Models
+
+### Recurrent Models (RNN/LSTM/GRU)
+- Inherently sequential computation → limited parallelization
+- O(n) sequential operations for long-range dependencies
+
+### Convolutional Models
+- Fixed kernel size restricts context → requires multiple layers
+- Logarithmic path length for distant connections
+
+---
+
+## Key Innovation: Self-Attention Mechanism
+
+- Connects all positions with constant operations
+- Enables parallel computation across sequence
+- Directly models long-range dependencies
+- More efficient than RNN/CNN for typical sequence lengths
+
+---
+
+## Transformer Model Architecture
+
+<img src="/paper_figure_1_transformer.png" class="h-[560px] w-full object-contain" />
+
+Encoder-decoder structure with stacked self-attention and feed-forward layers.
+
+---
+
+## Encoder Architecture
+
+- Stack of 6 identical layers
+- Each layer has two sub-layers:
+  1. Multi-head self-attention mechanism
+  2. Position-wise feed-forward network
+- Residual connections + layer normalization
+- Output dimension: dmodel = 512
+
+---
+
+## Decoder Architecture
+
+- Stack of 6 identical layers
+- Three sub-layers per layer:
+  1. Masked multi-head self-attention (prevents future positions)
+  2. Encoder-decoder attention (queries from decoder, keys/values from encoder)
+  3. Position-wise feed-forward network
+- Residual connections + layer normalization
+
+---
+
+## Attention Mechanisms
+
+### Scaled Dot-Product Attention
+$$\text{Attention}(Q, K, V) = \text{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right)V$$
+- Scaling avoids gradient vanishing for large dk
+
+### Multi-Head Attention
+$$\text{MultiHead}(Q, K, V) = \text{Concat}(\text{head}_1,...,\text{head}_h)W^O$$
+- h=8 parallel heads, dk=dv=64 → captures diverse patterns
+
+---
+
+## Positional Encoding
+
+- Injects sequence order information (no recurrence/convolution)
+- Added to input embeddings (same dmodel dimension)
+- Uses sine/cosine functions with varying frequencies:
+  $$PE_{(pos,2i)} = \sin(pos/10000^{2i/d_{\text{model}}})$$
+  $$PE_{(pos,2i+1)} = \cos(pos/10000^{2i/d_{\text{model}}})$$
+- Enables learning of relative position relationships
+
+---
+
+## Why Self-Attention?
+
+| Aspect               | Self-Attention | Recurrent | Convolutional |
+|----------------------|----------------|-----------|---------------|
+| Complexity           | O(n²·d)        | O(n·d²)   | O(k·n·d²)     |
+| Parallelization      | O(1)           | O(n)      | O(1)          |
+| Long-range path length | O(1)         | O(n)      | O(logk(n))    |
+
+- Superior parallelization and dependency modeling
+
+---
+
+## Training Setup
+
+### Data & Hardware
+- WMT 2014 EN-DE (4.5M) and EN-FR (36M) sentence pairs
+- 8 NVIDIA P100 GPUs, 12h (base)/3.5d (big model) training
+
+### Optimization
+- Adam (β1=0.9, β2=0.98, ϵ=10⁻⁹) with linear warmup (4000 steps) + inverse square root decay
+- Regularization: residual dropout (0.1), label smoothing (0.1)
+
+---
+
+## Machine Translation Results
+
+| Model | EN-DE BLEU | EN-FR BLEU | Training Cost (FLOPs) |
+|-------|------------|------------|-----------------------|
+| GNMT + RL Ensemble | 26.30 | 41.16 | 1.8×10²⁰ / 1.1×10²¹ |
+| ConvS2S Ensemble | 26.36 | 41.29 | 7.7×10¹⁹ / 1.2×10²¹ |
+| **Transformer (big)** | **28.4** | **41.8** | **2.3×10¹⁹** |
+
+- New state-of-the-art with 4× lower training cost
+
+---
+
+## Model Ablations & Generalization
+
+### Key Ablations (EN-DE Dev Set)
+| Variation | Dev BLEU | Insight |
+|-----------|----------|---------|
+| Single head | 24.9 | Multi-head critical |
+| No dropout | 25.3 | Regularization needed |
+| Learned pos encoding | 25.7 | Sinusoidal ≈ learned |
+
+### Constituency Parsing
+- 91.3 F1 (WSJ only) vs. 91.7 (state-of-the-art)
+- 92.7 F1 (semi-supervised) → strong generalization
+
+---
+
+## Attention Visualization: Long-Distance Dependencies
+
+<img src="/paper_figure_3_long_distance.png" class="h-[650px] w-full object-contain" />
+
+Encoder self-attention (layer 5) showing how "making" attends to distant words to complete the phrase "making...more difficult".
+
+---
+
+## Attention Visualization: Anaphora Resolution
+
+<img src="/paper_figure_4_anaphora.png" class="h-[650px] w-full object-contain" />
+
+Attention heads resolving the pronoun "its" to its referent "The Law" with sharp attention focusing.
+
+---
+
+## Limitations
+
+- Quadratic complexity in sequence length
+- Less efficient for very long sequences
+- Still sequential in generation process
+
+---
+
+## Future Work
+
+- Local/restricted attention mechanisms
+- Extension to other modalities (images, audio)
+- Non-sequential generation approaches
+- Efficient handling of large inputs/outputs
+
+---
+
+## Conclusion
+
+- Transformer replaces recurrence/convolution with self-attention
+- Sets new state-of-the-art in machine translation
+- Faster training via parallelization
+- Generalizes well to diverse sequence tasks
+- Foundation for modern attention-based NLP models

chapter5/paper-to-ppt/validation/runs/exp5-4-real-pdf-both-20260730-v5/source/source_visuals/manifest.json

@@ -0,0 +1,62 @@
+{
+  "paper": {
+    "title": "Attention Is All You Need",
+    "authors": "Ashish Vaswani et al.",
+    "arxiv_id": "1706.03762",
+    "pdf_url": "https://arxiv.org/pdf/1706.03762",
+    "pdf_sha256": "bdfaa68d8984f0dc02beaca527b76f207d99b666d31d1da728ee0728182df697",
+    "observed_pdf_sha256": "bdfaa68d8984f0dc02beaca527b76f207d99b666d31d1da728ee0728182df697"
+  },
+  "paper_text": {
+    "path": "/Users/boj/book/ai-agent-book/chapter5/paper-to-ppt/validation/runs/exp5-4-real-pdf-both-20260730-v5/source/paper_text.md",
+    "sha256": "4da5f2a1da38bad8149832b5ada3c94f0619188db8fed5cdca8ff63405587eb8",
+    "characters": 39819
+  },
+  "visuals": [
+    {
+      "filename": "paper_figure_1_transformer.png",
+      "pdf_page": 3,
+      "source_label": "Figure 1",
+      "caption": "The Transformer model architecture.",
+      "rect": [
+        92,
+        28,
+        520,
+        425
+      ],
+      "sha256": "1729d86922c9f97703af178fb52e4f876cc1fce1f67572730580ef16bd659bd5",
+      "bytes": 100778,
+      "public_copy_sha256": "1729d86922c9f97703af178fb52e4f876cc1fce1f67572730580ef16bd659bd5"
+    },
+    {
+      "filename": "paper_figure_3_long_distance.png",
+      "pdf_page": 13,
+      "source_label": "Figure 3",
+      "caption": "Encoder self-attention following long-distance dependencies.",
+      "rect": [
+        92,
+        55,
+        525,
+        365
+      ],
+      "sha256": "da9498d3022d1e0fc4d5da2bf59c10c9dde68dda471586019afa126db5674322",
+      "bytes": 93837,
+      "public_copy_sha256": "da9498d3022d1e0fc4d5da2bf59c10c9dde68dda471586019afa126db5674322"
+    },
+    {
+      "filename": "paper_figure_4_anaphora.png",
+      "pdf_page": 14,
+      "source_label": "Figure 4",
+      "caption": "Attention heads involved in anaphora resolution.",
+      "rect": [
+        92,
+        135,
+        525,
+        600
+      ],
+      "sha256": "8dee94773a369dccb1d766c218cc0ac16629a6baf79d15a313d13399d9064407",
+      "bytes": 242229,
+      "public_copy_sha256": "8dee94773a369dccb1d766c218cc0ac16629a6baf79d15a313d13399d9064407"
+    }
+  ]
+}