Linear Artificial Tomography maps Qwen 0.6B emotions

The author introduces a method called Linear Artificial Tomography scans to visualize and analyze emotions and behaviors in language models. Using contrastive prompt pairs and principal component analysis, the technique maps how emotional representations develop layer by layer within models like Qwen 0.6B. The author suggests that applying correction vectors at approximately 75 percent of the model's layers can help steer emotional outputs without harming overall capabilities. Open article →