Refusal Directions in Instruction-Tuned Models

This project explores how refusal behavior is represented inside instruction-tuned language models with the central question being: is refusal encoded as a single linear direction in the residual stream, or as a higher-dimensional structure?

Across models, both patterns appear.

---

Detailed Writeups

• Qwen-1.8B-Chat
• LLaMA-3.2-1B-Instruct
• Refusal Structure Across Models

These pages contain:

• Refusal vector extraction
• Cross-layer cosine analysis
• Runtime ablation experiments
• Subspace interventions (where applicable)
• Offline weight orthogonalization
• Benchmark evaluation