paper
referenced-only
2024
paper:2024-activationActivation addition: Steering language models without optimization
Similar preprints — Semantic Scholar
Cited by (2)
- Belief Dynamics Reveal the Dual Nature of In-Context Learning and Activation Steering
In-context learning and activation steering are the same operation on a latent-concept belief: context adds log-likelihood evidence while a steering vector adds a constant shift to the log prior odds,
- Facet-Level Persona Control by Trait-Activated Routing with Contrastive SAE for Role-Playing LLMs
Facet-level personality control in role-playing LLMs is substantially improved by injecting contrastively trained sparse autoencoder (SAE) control vectors into mid-residual layers, with the CV-SAE+Pro