New Interface Design Lets Users Control Chatbot Personality Without Typing Prompts

A team of researchers has proposed a new approach to personalizing large language models that sidesteps the friction of writing detailed text instructions. Instead of asking users to articulate their preferences in natural language, the system uses direct manipulation—leveraging a simple control mechanism to adjust how strongly a particular trait or behavior is expressed in the chatbot's responses.

How Activation Steering Works#

The core innovation centers on activation steering, a technique that employs a linear scalar to dial preference intensity up or down. Rather than requiring users to prompt their way to the desired behavior, this method lets them adjust a single factor to control output expression. The researchers first validated that this approach is computationally practical for fine-grained preference control, then designed three prototype interfaces to make the mechanism accessible to end users.

Three Interface Designs Tested#

The designs differed along two key dimensions:

• Agency: Whether users took the lead in adjusting settings or the system made suggestions
• Fluidity: Whether controls remained fixed or adapted based on user interaction

The research team conducted a within-subjects study with 14 participants performing cold-start personalization tasks—situations where the chatbot has no prior knowledge of the user's preferences.

What the Study Revealed#

Results suggested that steerable chatbots could align more closely with user preferences than prompting alone. However, the findings also highlighted diversity in user attitudes. Participants expressed varying priorities around control, persistence (whether settings stick across conversations), and transparency about how the system works internally. This diversity implies that no single interface design will suit all users, and future personalization systems may need to offer flexibility in how users interact with these controls.