SafeSteer uses category-specific activation vectors to steer LLMs toward safe, on-topic, non-refusing responses at inference time.
A group of people using the name of Allah to bullyand harass a group of people based on their religion online.3
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
SafeSteer: Interpretable Safety Steering with Refusal-Evasion in LLMs
SafeSteer uses category-specific activation vectors to steer LLMs toward safe, on-topic, non-refusing responses at inference time.