← AI Terminology
Activation Steering
Activation steering edits model behaviour at inference by adding direction vectors into activations (or similarly soft-controlling features) without changing weights.
Used for style, truthfulness experiments, and safety research.
Used for style, truthfulness experiments, and safety research.
Why It Matters in AI
Fine-tuning is heavy; steering probes whether concepts are linearly readable and controllable. It connects interp to practical behaviour modulation — with dual-use concerns.
Key Points
| Aspect | Description |
|---|---|
| Use | Tone, refusal, sycophancy studies |
| Risk | Bypass refusals if misused |
| Limits | Side effects; brittle across prompts |
| Method | Add α·v to activations at chosen layers |
| Related | Representation engineering, SAE features |
| Vectors | Contrastive pairs; SAE features; CCA directions |
Simple Analogy
Gently nudging a conversation by whispering a mood into someone’s ear mid-sentence — same brain weights, shifted trajectory.
Common Usage Examples
- Contrast “love vs hate” residual directions
- Steer with SAE feature vectors
- Measure utility vs capability side effects
- Layer sweeps for best control
Summary
In short: Activation steering adds control vectors into running activations to shift behaviour without retraining weights.