analiza aiqbench
This research introduces REINS, a novel steering technique using Sparse Autoencoders (SAEs) to robustly suppress harmful outputs in large language models without retraining. By simultaneously inhibiting dangerous continuation features and enhancing refusal mechanisms, the method addresses critical vulnerabilities where complex adversarial wrappers previously undermined safety controls. The approach demonstrates significant improvements in reliable refusal rates while maintaining general model capabilities, offering a lightweight path for safer AI deployment.
REINS: Refusal-Enhanced Inhibitory Steering with Sparse Autoencoder Features
arXiv:2608.28233v1 Announce Type: new Abstract: 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. Existing single direction SAE steering methods do not reliably produce refusals on harmful prompts, suggesting that refusal enhancement alone can be too weak when the harmful continuation path remains active. This motivates us to propose Refusal-Enhanced INhibitory Steering (REINS), which suppresses harmful continuation features and enhances safe refusal features in the same SAE feature space. Experiments on GUISE and other datasets show that prior methods either intervene too weakly or achieve only apparent safety through collapse, while REINS substantially reduces harmful responses, markedly improves safe refusals and largely preserves general capabilities.
Abstract: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. Existing single direction SAE steering methods do not reliably produce refusals on harmful prompts, suggesting that refusal enhancement alone can be too weak when the harmful continuation path remains active. This motivates us to propose Refusal-Enhanced INhibitory Steering (REINS), which suppresses harmful continuation features and enhances safe refusal features in the same SAE feature space. Experiments on GUISE and other datasets show that prior methods either intervene too weakly or achieve only apparent safety through collapse, while REINS substantially reduces harmful responses, markedly improves safe refusals and largely preserves general capabilities.
Artificial Intelligence (cs.AI)
arXiv:2608.28233 [cs.AI]
(or arXiv:2608.28233v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2608.28233 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 264 cuvinte
Subiecte conexe
CircuitSteer: Geometrically Aligned Multi-Layer Steering via Sparse Autoencoder Circuits
arXiv · 7 aug. 2026
There Is More to Refusal in Large Language Models than a Single Direction
arXiv · 16 sept. 2026
Latent Space Refusal Anchoring for Low-Resource African Languages: Mechanistic Safety Recovery Without Retraining
arXiv · 21 aug. 2026
EmoRSS: Mitigating Emotion-Induced Over-Refusal in Large Language Models
arXiv · 7 oct. 2026