Skip to content

Activation-steering methods

Add or remove safety-related directions in the residual stream. Most methods wrap a frozen model with forward hooks that modify hidden states at inference time; a few (CircuitBreaker, CircuitBreakerRR) are training-time methods grouped here because they also target hidden representations.

Method Mechanism Reference
RefusalDirectionTrainer Single direction steer/ablate Arditi et al., NeurIPS 2024
CAATrainer Contrastive activation addition Panickssery et al., 2023
AdaSteerTrainer Adaptive two-direction steering Zhao et al., EMNLP 2025
SafeSteerTrainer Category-routed steering 2025
AlphaSteerTrainer Closed-form ridge regression 2025
SafeSwitchTrainer Two-stage prober + refusal head Han et al., EMNLP 2025
SCANSTrainer Adaptive-sign refusal steering Cao et al., AAAI 2025
STATrainer SAE latent steering Wang et al., ACL 2025
CircuitBreakerTrainer Representation rerouting Zou et al., NeurIPS 2024
CircuitBreakerRRTrainer Dual-purpose RR Zou et al., NeurIPS 2024
RepBendTrainer Representation bending Yousefpour et al., ACL 2025
LinearProbeGuardTrainer Probe-based guarding
CASTTrainer Conditional activation steering Wu et al., ICLR 2025
TARSteerTrainer Tamper-resistant safeguards Tamir et al., 2024
RRFAEnsembleTrainer RR for agentic safety Zhou et al., 2025