ALTSTEER: Selective Safety Steering for Moving Beyond Hard Refusals to Constructive Alternatives

Safety alignment is essential for deploying large language models (LLMs), requiring systems to prevent harmful compliance while preserving helpfulness on benign requests. Activation steering offers a training-free inference-time approach to safety control, but effective safety steering requires addressing two coupled questions: **when to intervene** and **how generation should be shaped after intervention**. Existing safety steering methods remain limited along both dimensions: their triggering mechanisms can be unstable across domains, and refusal-oriented steering often yields rigid refusals rather than constructive safe guidance. To address these limitations, the authors propose ALTSTEER, an inference-time framework that couples selective intervention with refusal-anchored constructive redirection within a single inference pass. The work targets a practical gap: moving beyond hard refusals to constructive alternatives while maintaining benign utility.

Evaluations on Llama-3.1 and Qwen2.5 show that ALTSTEER preserves benign utility while improving constructive safe-completion behavior, especially on models that otherwise tend to produce short refusals for harmful requests. The authors report that ALTSTEER achieves a better balance between safety and helpfulness compared to baseline steering methods. Specifically, on models prone to short refusals, ALTSTEER increases the rate of constructive safe completions without significantly degrading performance on benign tasks. The abstract does not provide specific numerical values, but the qualitative results indicate consistent improvements across the evaluated models.