Abstract
Feature steering enables users to strengthen or weaken concepts (features) that an LLM has already learned in real Ime. It has emerged as an alterna Ive to prompIng, allowing conInuous rather than discrete control over model behavior. However, given the technical knowledge required to extract and then steer on features, this process is mostly limited to technical researchers. We seek to build a graphical user interface that makes steering accessible to the layperson. ExisIng interfaces prioriIze flexibility over intuiIon, overwhelming users with the sheer number of opIons for what to prompt or what features to steer on. We conducted a user-centered design process guided by two phases of user research to develop our GUI. Through user tesIng, we found that the most valuable use case for feature steering for the layperson is creaIng model personas for specific use cases. We developed our interface with this use case in mind, and engaged in formaIve and summaIve tesIng to evaluate design decisions related to represen Ing features, building steering intuiIon, and reducing the blank slate problem. Our interface, built on Goodfire's Ember API for LLaMA 3.1 8b, simplifies steering controls and provides clear response comparisons to help users build intuiIon about how steering affects outputs. This work demonstrates how interface design can make powerful but complex interpret ability tools more accessible, allowing everyday users to meaningfully shape model behavior through steering. Code is available at h[ps://github.com/acyhuang/steering-interface.
Keywords
Feature steering; Human-centered AI; Interface design; Large language models
DOI
https://doi.org/10.21606/iasdr.2025.601
Citation
Huang, A., Lim, Y.,and Angeles, L.(2025) Designing Intuitive Interfaces for Feature Steering of LLMs, in Chang, C.-Y., and Hsu, Y. (eds.), IASDR 2025: Design Next, 02-05 December, Taiwan. https://doi.org/10.21606/iasdr.2025.601
Creative Commons License
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License
Conference Track
Track 4 - Human-Centered AI
Designing Intuitive Interfaces for Feature Steering of LLMs
Feature steering enables users to strengthen or weaken concepts (features) that an LLM has already learned in real Ime. It has emerged as an alterna Ive to prompIng, allowing conInuous rather than discrete control over model behavior. However, given the technical knowledge required to extract and then steer on features, this process is mostly limited to technical researchers. We seek to build a graphical user interface that makes steering accessible to the layperson. ExisIng interfaces prioriIze flexibility over intuiIon, overwhelming users with the sheer number of opIons for what to prompt or what features to steer on. We conducted a user-centered design process guided by two phases of user research to develop our GUI. Through user tesIng, we found that the most valuable use case for feature steering for the layperson is creaIng model personas for specific use cases. We developed our interface with this use case in mind, and engaged in formaIve and summaIve tesIng to evaluate design decisions related to represen Ing features, building steering intuiIon, and reducing the blank slate problem. Our interface, built on Goodfire's Ember API for LLaMA 3.1 8b, simplifies steering controls and provides clear response comparisons to help users build intuiIon about how steering affects outputs. This work demonstrates how interface design can make powerful but complex interpret ability tools more accessible, allowing everyday users to meaningfully shape model behavior through steering. Code is available at h[ps://github.com/acyhuang/steering-interface.