Research question
Can adding vectors to activations steer generation without training?
Key findings
- Activation addition with contrastive prompt pairs steers sentiment and topic at inference time - Preserves capabilities on unrelated inputs
Steering Language Models With Activation Engineering
Activation addition (ActAdd), the simple inference-time steering technique that seeded the steering-vector literature.
Contribute
Spot an error, or know a record we should add? Members can propose additions and submit corrections.