SOURCE-LINKED INTELLIGENCE

REINS: Refusal-Enhanced Inhibitory Steering with Sparse Autoencoder Features

arXiv · AI, language, vision and robotics · article · Aug 28, 2026 · UTC

Steering with Sparse Autoencoders (SAEs) offers a lightweight inference-time path for adapting the behavior of large language models without retraining. By exposing sparse and interpretable features, SAE steering provides a promising interface for safety control that guides harmful continuations toward refusal. However, we observe that complex wrappers can still undermine existing SAE steering methods on harmful prompts. To evaluate this failure mode systematically, we construct Generalized Undercover Instruction Safety Evaluation (GUISE), a dataset of harmful prompts with complex wrappers. Ex

Read original source ↗ Open in workspace

recordType

paper

region

Global