I am a visiting research scientist at Google PAIR (People + AI Research). I generally research Human-AI Interaction, with a focus on human interaction with LLMs and making LLMs easier to use.
I received my PhD in computer science (specialized in HCI, AI, and Human-AI Interaction) from Columbia University, where I was lucky to be advised by Prof. Lydia B. Chilton. I was also fortunate to have research internships at Spotify, Adobe, and IBM.
I'm on the job market for research scientist / engineer roles!
Feel free to reach out to me at: savvas@cs.columbia.edu
ConstitutionalExperts: Training a Mixture of Principle-based Prompts
ConstitutionalExperts automatically improves LLM prompts via fine-grained principle edits. Unlike prior methods that optimize the prompt as a single entity, our method incrementally improves the prompt by surgically editing individual principles.
In submission at ACL 2024, [pdf]ConstitutionMaker: Interactively Critiquing Large Language Models by Converting Feedback into Principles
ConstitutionMaker enables users to interactively improve LLM prompts with natural language feedback. Users can critique, kudos, or rewrite the LLM's output, and this feedback is then used to automatically update the prompt with a specific principle.
IUI 2024, [pdf]PromptInfuser: How Tightly Coupling AI and UI Design Impacts Designers' Workflows
PromptInfuser is a Figma plugin that enables users to create functional medium-fidelity prototypes, by connecting the UI mockup to LLM prompts. In a study with 14 designers, we found that PromptInfuser encouraged iteration over prompt and UI together, which helped designers identify UI and prompt incompatibilities and reflect upon their total solution.
In submission at DIS 2024, [pdf]AngleKindling: Supporting Journalistic Angle Ideation with Large Language Models
AngleKindling is an interactive web application which employs the common sense reasoning of LLMs to help journalists explore angles (e.g. controversies) for reporting on a press release. In a study with 12 professional journalists, we show that participants found AngleKindling signifcantly more helpful and less mentally demanding to use for brainstorming ideas, compared to a prior journalistic angle ideation tool.
CHI 2023, [pdf]