# Technical Document Extraction: Chain-of-Thought (CoT) and Chain of Continuous Thought (CoCoNUT) Diagrams
## Diagram Overview
The image contains two side-by-side diagrams comparing **Chain-of-Thought (CoT)** and **Chain of Continuous Thought (CoCoNUT)** architectures in large language models. Both diagrams illustrate token processing flows and hidden state utilization.
---
### **Left Diagram: Chain-of-Thought (CoT)**
#### Components and Flow
1. **Input Section**:
- **Input Token**: `[Question]` (yellow square).
- **Input Embedding**: Represented by a sequence of yellow squares labeled `x_i`, `x_i+1`, `x_i+2`, ..., `x_i+j`.
2. **Processing**:
- **Large Language Model**: Central gray rectangle labeled "Large Language Model."
- **Hidden States**: Purple squares labeled "last hidden state" (e.g., `x_i`, `x_i+1`, ..., `x_i+j`).
3. **Output Section**:
- **Output Token (Sampling)**: Purple squares labeled `x_i`, `x_i+1`, ..., `x_i+j`.
- **Final Output**: `[Answer]` (purple square).
#### Legend and Spatial Grounding
- **Legend**: Located on the right side of the diagram.
- **Purple**: Last hidden states.
- **Yellow**: Input embeddings.
- **Spatial Flow**: Tokens flow left-to-right through the model, with hidden states feeding into output token sampling.
---
### **Right Diagram: Chain of Continuous Thought (CoCoNUT)**
#### Components and Flow
1. **Input Section**:
- **Input Token**: `[Question]` (yellow square).
- **Special Tokens**: `<bot>` and `<eot>` (yellow squares).
- **Input Embedding**: Represented by a sequence of yellow squares.
2. **Processing**:
- **Large Language Model**: Central gray rectangle labeled "Large Language Model."
- **Hidden States**: Orange squares labeled "last hidden states" (e.g., `x_i`, `x_i+1`, ..., `x_i+j`).
- **Feedback Loop**: Hidden states are reused as input embeddings for subsequent steps (indicated by bidirectional arrows).
3. **Output Section**:
- **Output Token (Sampling)**: Orange squares labeled `x_i`, `x_i+1`, ..., `x_i+j`.
- **Final Output**: `[Answer]` (purple square).
#### Legend and Spatial Grounding
- **Legend**: Located on the right side of the diagram.
- **Orange**: Last hidden states.
- **Yellow**: Input embeddings.
- **Spatial Flow**: Hidden states are recycled as input embeddings, creating a continuous loop within the model.
---
### Key Differences
| Feature | CoT | CoCoNUT |
|------------------------|------------------------------|------------------------------|
| **Hidden State Use** | Linear progression | Recycled as input embeddings |
| **Special Tokens** | None | `<bot>`, `<eot>` |
| **Feedback Mechanism** | None | Explicit feedback loop |
---
### Notes
- **No Numerical Data**: Both diagrams are conceptual and do not include numerical values or statistical trends.
- **Color Consistency**: Legend colors (purple, yellow, orange) are strictly adhered to in both diagrams.
- **Textual Labels**: All labels (e.g., `[Question]`, `[Answer]`, `<bot>`) are transcribed verbatim.
This extraction provides a complete technical breakdown of the diagrams, ensuring no component or label is omitted.
