一作ACL 2026Findings· Findings of ACL
Analysing the Safety Pitfalls of Steering Vectors
Yuxiao Li · Alina Fastowski · Efstratios Zaradoukas · Bardh Prenkaj · Gjergji Kasneci
摘要
Activation steering has emerged as a powerful tool to shape LLM behaviour without the need for weight updates. While its inherent brittleness and unreliability are well-documented, its safety implications remain underexplored. In this work, we present a systematic safety audit of steering vectors...