research

Analytics Needs of Video Game Streamers

Conducted interview studies with streamers from Twitch and Mixer to understand the information gap in their streaming process and identify their analytics needs.

Human-LM Alignment

Tested popular large language models on high-stakes decision making tasks like recidivism prediction, analyzing their agreement with human predictions and existing models. Found that providing different types of information, including distractor information like photos, can significantly influence LM decisions and that current LMs are not the right tools for these types of tasks.

Advancing Human-AI Complementarity

Examined how users with varying levels of expertise interact with AI recommendations in a blood vessel labeling task. Found that user base expertise, complementary AI tuning to address users' strengths and weaknesses, and users' perception of AI performance all significantly impact Human-AI team performance.

Race Presentation on Human Judgement of Recidivism

Explored how the saliency of race information presentation influences users' recidivism judgements. Found that adding photos to race labels had a significant impact on predictions for users who identified as female, with black defendants being less likely to be predicted to recidivate compared to white defendants.

Perspectives

Designed and evaluated a novel hybrid meeting system that digitally composites everyone into a virtual room where each person has a unique but spatially consistent viewpoint. Found that this approach supports natural interactions, creates a strong sense of co-presence, and reduces cognitive load compared to commercial alternatives like Gallery, Together Mode, and Front Row.