Nonlinear steering improves control over AI language model behavior

Original: Beyond Linear Activation Steering: Invertible Latent Transformations for Controlling LLM Behavior

Source: arxiv.org ↗