Agentic Harnesses for AI4S
I study how models can retain and revise scientific work through empirical feedback, starting with SimAuthor ↗.
Currently working on
Learning Without Updating Weights
I am exploring how audio skills can improve through feedback and text optimization, without updating model weights.
AI for Health Signals & Care
I work with hospitals on health signals, including clinical audio in StethoLM ↗ and signal simulation in SimAuthor ↗.
Archive
Language, Bias & Human Preferences
Two earlier papers examine negation bias ↗ in language models and how empathy and demographics ↗ shape people's preferences for LLM responses.