activation space
Activation space refers to the multi-dimensional space defined by the activations of neurons in a neural network during processing. Analyzing activation space can provide insights into model behavior, interpretability, and decision-making processes.
- A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoders
- A is for Absorption: Studying Feature Splitting and Absorption in Sparse Autoencoders
- Activation-Informed Merging of Large Language Models
- Angular Steering: Behavior Control via Rotation in Activation Space
- Mitigating Overthinking in Large Reasoning Models via Manifold Steering
- Refusal Direction is Universal Across Safety-Aligned Languages