Modify activations, not weights
Steering vectors operate on model activations during inference. No retraining, no fine-tuning, no weight modification. Just precise behavioral adjustments in real-time.
Example: Adjusting Output Style
Pre-trained Vector Library
50+ ready-to-use vectors for common behaviors: formality, conciseness, helpfulness, creativity, assertiveness, and more. Start controlling behavior immediately.
Custom Vector Training
Train custom steering vectors for your specific use cases. Provide contrastive examples of desired vs. undesired behavior, and Steer learns the activation direction.
Composable Vectors
Combine multiple steering vectors for nuanced control. Layer formality with conciseness, or creativity with accuracy. Adjust strength of each independently.
Real-time Adjustment
Change steering parameters via API without redeployment. Adjust vector strength on the fly. Respond to user feedback or business requirements instantly.
Research Foundation: Activation Steering
Steering vectors work by adding direction vectors to model activations at specific layers. Research shows this can reliably modify behavior without degrading overall model quality.
Open source: rotalabs-steer toolkit available at rotalabs.ai. Train your own vectors, verify our methods.