🤖 AI

Steerable Chatbots: Exploring Personalization Control Interfaces via LLM Activation Steering

This paper introduces "steerable chatbots" as a novel personalization paradigm that allows users to directly control LLM output via activation steering rather than prompting, and evaluates three interface designs through a user study to demonstrate their potential for better preference alignment while highlighting diverse user values regarding control and transparency.

Original authors: Jessica Y. Bo, Tianyu Xu, Ishan Chatterjee, Katrina Passarella-Ward, Achin Kulshrestha, D Shin

Published 2026-07-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jessica Y. Bo, Tianyu Xu, Ishan Chatterjee, Katrina Passarella-Ward, Achin Kulshrestha, D Shin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are talking to a super-smart robot that knows everything in the world. Usually, if you want it to act a certain way—like being super fancy or super cheap—you have to explain it using words. You might say, "I want a fancy dinner," or "I'm on a tight budget." But sometimes, it's hard to find the right words, or the robot just doesn't quite "get" what you mean, especially when you first start talking to it. This paper is about a new way to talk to these robots, not just with words, but with a "volume knob" for their personality.

The scientists behind this study are working in a field called Human-Computer Interaction, which is basically the study of how humans and machines get along. They are using a technique called "activation steering." Think of a large language model (the brain of the chatbot) like a giant, complex radio. Usually, you change the station by typing a request. But activation steering is like having a dial that directly turns up or down specific "frequencies" inside the robot's brain. If you turn the "Luxury" dial up, the robot starts thinking about expensive things. If you turn the "Budget" dial up, it thinks about saving money. The big question the researchers asked was: If we give regular people this special dial, how should it look? Should they slide it back and forth? Should they set it once and forget it? Or should the robot figure it out on its own?

This paper introduces a new idea called "Steerable Chatbots." Instead of asking you to write a perfect paragraph to tell the robot what you like, the researchers built interfaces that let you control the robot's personality directly using a simple number or slider. They tested three different ways to do this. The first, called SELECT, is like a volume slider on your phone; you can drag it up or down right while you are chatting to instantly change the robot's vibe. The second, CALIBRATE, is like setting up a new pair of glasses; before you start talking, the robot shows you two different answers, and you pick the one you like better, helping it find the perfect setting before the conversation even begins. The third, LEARN, is like a mind-reading friend; the robot listens to what you say and guesses your preferences, adjusting its settings automatically in the background without you touching anything.

The researchers ran computer tests first to make sure the "dials" actually worked. They found that turning the dial did indeed change the robot's answers, making it talk about budget restaurants when turned one way and fancy ones when turned the other. Then, they invited 14 real people to try out these three new chatbots. The results were interesting: all three ways of using the dial helped the robot understand the users better than just asking the robot to guess based on words alone. However, people didn't all like the same way of controlling it. Some people loved having the slider (SELECT) because they felt totally in charge. Others preferred the robot to figure it out (LEARN) because it felt more natural and less work. A few liked the setup step (CALIBRATE) because it felt precise, like getting an eye exam.

The study suggests that giving people a direct "knob" to control a chatbot's personality is a powerful tool, especially when you first start using it and don't have a long history of chats to teach the robot. But there isn't one perfect design for everyone. Some people want to be the captain of the ship, while others want the ship to sail itself. The paper concludes that the future of talking to AI might not just be about better typing, but about giving us different kinds of controls—sliders, setup wizards, or auto-pilots—so we can choose how much we want to steer the conversation ourselves.

Technical Summary: Steerable Chatbots via LLM Activation Steering

Problem Statement
Current paradigms for personalizing Large Language Model (LLM) responses rely heavily on users articulating preferences through natural language prompting. This approach presents two significant barriers: (1) the "cold-start" problem, where users lack conversational history for the model to retrieve preferences, and (2) the cognitive burden of structuring complex preferences into effective prompts, a skill many lay users lack. Existing technical solutions like retrieval-augmented generation (RAG) or model fine-tuning are computationally intensive and require substantial user data, making them unsuitable for immediate, granular personalization. The authors propose a shift from prompting-based specification to direct manipulation of model behavior to align LLM outputs with latent user preferences.

Methodology
The paper introduces "Steerable Chatbots," a paradigm where users manipulate a linear scalar factor to control the strength and direction of specific preference dimensions within an LLM's output. This is achieved through activation steering, a technique that injects pre-computed "steering vectors" into the model's residual stream during inference.

Technical Implementation:
- Steering Vectors: The authors compute vectors by contrasting desirable and undesirable representations of a behavior (e.g., "luxury" vs. "budget") using layer-wise linear probes trained on positive and negative data samples.
- Steering Mechanism: During inference, the steering vector ( $v$ ) is added to the intermediate representations ( $h$ ) of the LLM, scaled by a user-defined strength factor ( $d$ ): $h_{steered} \leftarrow h + d \cdot v$ .
- Evaluation Metrics: Preference expression is quantified using the relative cosine similarity between the LLM's output embeddings and reference datasets (Yelp reviews) representing specific preference traits (Cost, Ambiance, Age, Time, Cuisine).
Interface Designs:
The authors explore three distinct interface designs mapped across two axes: Agency (User-led vs. System-driven) and Fluidity (Static vs. Adaptive):
- SELECT (High Agency, Medium Fluidity): Users directly manipulate a slider (-100 to 100) in real-time to adjust the steering factor. This allows for immediate, explicit control over preference dimensions.
- CALIBRATE (Medium Agency, Low Fluidity): Before the conversation, users undergo a calibration process where they iteratively select preferred outputs from pairs of responses generated with different steering strengths. The system converges on a static preference value for the session.
- LEARN (Low Agency, High Fluidity): The system implicitly infers user preferences from conversational sentiment and updates the steering factor in the background without explicit user input, displaying the inferred strength for transparency.
Experimental Setup:
- Computational Experiments: Validated across five open-source LLMs (ranging from 1.6B to 9B parameters) to ensure steering reliability (E1) and its composability with user prompting (E2).
- User Study: A within-subjects study ( $n=14$ ) involving cold-start lifestyle planning tasks (gift shopping, travel, dining, recipes). Participants interacted with the three steering interfaces and a prompting-only baseline (PROMPT).

Key Results

Computational Viability: Activation steering demonstrated a strong, near-linear correlation between the steering strength and the expressed preference across all tested models. Crucially, the method was found to be compatible with user prompting; prompting could intensify or reduce the effect of steering, allowing for contextual variability within a steered preference profile.
Preference Alignment (U1):
- SELECT and CALIBRATE showed moderate, significant correlations ( $r=0.46$ and $r=0.59$ respectively) between the chatbot's output and the user's ground-truth preferences, significantly outperforming the prompting-only baseline ( $r=0.11$ ).
- LEARN showed a weaker correlation ( $r=0.26$ ), suggesting the sentiment-based update algorithm was less robust in capturing diverse user intentions in a short timeframe.
- Overall, steerable chatbots achieved a moderate correlation ( $r=0.44$ ) with latent preferences, validating the approach for cold-start scenarios.
Subjective Perceptions (U2):
- There was no significant difference in "Likelihood to Use," "Satisfaction," or "Perceived Control" across the interfaces when compared directly.
- However, all steerable interfaces were rated significantly higher than the baseline in Perceived Persistency, indicating users felt their preferences were better remembered across sessions.
- Heterogeneity of Values: Qualitative analysis revealed distinct user preferences. Some users valued the explicit control of SELECT, while others preferred the seamless, low-cognitive-load experience of LEARN. CALIBRATE was appreciated for its precision and time-saving nature but criticized for potential misalignment if the calibration step failed to capture nuanced preferences.

Significance and Claims
The paper claims to introduce a new paradigm for LLM personalization that shifts the burden of preference articulation from natural language prompting to direct manipulation of model internals. The authors posit that activation steering offers a resource-efficient alternative to fine-tuning, as it requires no retraining of model weights and operates with lightweight vectors.

The study reframes LLM personalization as an interaction design problem, demonstrating that exposing a linear steering factor can effectively align model outputs with latent user preferences in cold-start scenarios. The authors emphasize that there is no single "optimal" interface; rather, user values regarding agency (control vs. automation) and fluidity (static vs. adaptive) dictate which design is most suitable. They conclude that steerable chatbots offer a promising, computationally efficient mechanism for creating directly manipulatable, personalized AI assistants, though they acknowledge limitations regarding the fidelity of low-fidelity prototypes and the reduction of complex preferences to single dimensions.

Technical Summary: Steerable Chatbots via LLM Activation Steering

More like this