# Technical Document Extraction: Neural Network Layer Activation Analysis
## 1. Document Overview
This image contains a grid of 12 scatter plots representing the internal activations of a neural network (likely a Large Language Model) across various layers. The visualization tracks how the model differentiates between truthful statements, hallucinations, and intentional lies, and demonstrates the effect of "Honesty control" via a steering vector.
## 2. Component Isolation
### A. Header/Title Region
Each subplot is labeled with a specific layer number, indicating the depth within the model architecture.
* **Row 1:** Layer 2, Layer 4, Layer 7, Layer 10
* **Row 2:** Layer 11, Layer 12, Layer 13, Layer 14
* **Row 3:** Layer 16, Layer 20, Layer 26, Layer 31
### B. Main Chart Region (Data Visualization)
The charts use dimensionality reduction (likely PCA or t-SNE) to project high-dimensional activations into a 2D space.
#### Legend and Categories
Located at the bottom of the image:
* **Green Checkbox Icon:** `Truth`
* **Red 'X' Icon:** `Hallucination`
* **Orange Face Icon (🧐):** `Lie`
* **Black Arrow ($\leftarrow$):** `Steering vector`
* **Wrench Icon:** `Honesty control` (represented by light purple/grey clusters)
### C. Trend Verification and Data Analysis
| Layer | Visual Trend / Cluster Separation |
| :--- | :--- |
| **Layer 2** | High overlap. Truth, Hallucination, and Lie data points are mixed in a single central cloud. No clear distinction. |
| **Layer 4** | Initial separation. Lies (Orange) begin to cluster at the top right, while Truth and Hallucinations remain mixed at the bottom left. |
| **Layer 7** | Vertical separation. Truth/Hallucinations form a vertical band on the left; Lies form a distinct vertical cluster on the right. |
| **Layer 10** | Sharp separation. Truth/Hallucinations are tightly packed on the far left. Lies are on the far right. A new grey cluster (Honesty control) appears near the Lies. |
| **Layer 11** | Similar to Layer 10. A black arrow (Steering vector) is visible, pointing from the Honesty control cluster toward the Truth cluster. |
| **Layer 12** | The Honesty control cluster (grey) moves further left, away from the Lie cluster (orange), following the steering vector. |
| **Layer 13** | The Honesty control cluster is now positioned in the center of the vacuum between Lies and Truth. |
| **Layer 14** | The Honesty control cluster continues its trajectory toward the left (Truth/Hallucination) side. |
| **Layer 16** | Truth and Hallucinations begin to elongate into a diagonal "V" shape. The Honesty control cluster is closer to the Truth base. |
| **Layer 20** | Truth and Hallucinations show distinct "tails." The Honesty control cluster is overlapping with the lower section of the Truth/Hallucination distribution. |
| **Layer 26** | Truth (Green) and Hallucinations (Red) show significant divergence. Red points dominate the upper "tail," Green points dominate the lower "tail." |
| **Layer 31** | Final state. Clear tripartite separation: Hallucinations (Top Left), Truth (Bottom Left), and Lies (Far Right). The Honesty control cluster is integrated with Truth. |
## 3. Key Findings and Technical Insights
1. **Lie Detection:** The model distinguishes "Lies" (intentional falsehoods) very early (by Layer 4) and maintains a very high spatial distance between Lies and Truth throughout the remaining layers.
2. **Truth vs. Hallucination:** These categories are indistinguishable in early and middle layers. They only begin to spatially diverge in the very late stages of the model (Layer 26 and Layer 31), suggesting that "hallucination" is a more subtle internal state than "lying."
3. **Steering Vector Efficacy:** The "Honesty control" (grey points) demonstrates the application of a steering vector. The black arrows in Layers 10-14 show the vector's direction. The visualization proves that applying this vector successfully moves "Lie" activations toward the "Truth" manifold in the latent space.
4. **Manifold Geometry:** The data transitions from a disorganized cloud (Layer 2) to a linear separation (Layer 7) and finally to a complex, branched manifold (Layer 31) where different types of truthfulness occupy distinct "arms" of the distribution.
