Iterative Sparse Matrix Steering: Closed-Form Subspace Alignment

Figure: Visualizing the "Forbidden Manifold" alignment in 2D space.

---

📚 Select Language / Оберіть мову

Please select your preferred language to read the full paper and methodology: (Оберіть мову, щоб прочитати повну статтю та опис методології):

🇬🇧 English	🇺🇦 Українська
Read Full Paper	Читати статтю

---

🚀 Quick Start

Jump directly to the code implementation and experiments (Gemma3-1B):

➡️ Open Jupyter Notebook

Note: The notebook includes the full pipeline:

1. Data extraction
2. Ridge Regression training (Closed-Form)
3. Inference with MatrixSteeringHook
4. Visualization of the "Distillation Regime"

---

⚡ Key Features

• No Gradient Descent: Solves steering matrices analytically using Ridge Regression on CPU.
• Context-Aware: Unlike static vectors, matrix steering acts as an affine transformation ($h' = hW^T + b$), adapting to the token's context.
• Ontological Editing: Demonstrates how to robustly change model beliefs (e.g., "Moon is Cheese") using high-regularization distillation.