Fecha de publicación: -- Fuente: WIPO "tomato" Disclosure relates to controlling AI responses through preference-based activation steering. A prompt is received and associated with a preference represented as a specific position within a range between two opposing features associated with the preference (e.g., luxury vs. budget related to cost preference). This position, which indicates which feature is favored and by how much, is set via interactive user interface inputs across multiple modalities: direct selection of candidate preferences (select implementation), iterative calibration between model-generated responses (calibrate implementation), or learning from interactions (learn implementation). Based on the position determined from the input, a corresponding steering strength of the position is injected into the model's stream to influence response generation. This allows the AI to align its output with the favored features/attributes. The system enables dynamic adjustment of responses through multiple interactions, where each interaction updates the position to refine the personalization in real-time.