AI Safety & Ethics · Emerging · Advanced
Mechanistic Feature Steering
Also known as: Feature Steering, Concept Activation Vectors
A safety and control technique that modifies internal LLM hidden activations along specific sparse autoencoder feature directions.
What Mechanistic Feature Steering is
Mechanistic Feature Steering allows precise runtime control over model tone, refusal behavior, and topic focus without fine-tuning.
How it works
Adds positive or negative scalar vectors to activation tensors during forward passes.
Why it matters
Enables active defense against jailbreaks and real-time behavioral alignment.
Common uses
- →Jailbreak mitigation
- →Real-time model persona control
- →AI safety research
More in this collection
Browse all AI Concepts