Ruixin (Ray) Yang     杨瑞欣

I am an MSCS student at Georgia Tech, working with Prof. Alan Ritter. I received my BSc degree in Computer Science and Statistics from University of British Columbia in the beautiful Vancouver, Canada.

My research focuses on improving foundation models and agentic systems to make them more trustworthy and adaptive, enabling safe and effective human-AI collaboration. Currently, I am interested in:

(1) auditing, evaluating, and post-training for reliable agentic AI, especially in high-stakes, specialized domains (e.g., scientific and medical applications) and long-horizon tasks (where verification or oversight is difficult), and in scenarios requiring multimodal interaction or multi-agent coordination.

This includes improving both epistemic reliability (e.g., confidence calibration, recognizing knowledge gap, handling ambiguity) and behavioral reliability (e.g., mitigating risks of privacy leakage, tool misuse, and alignment failures).

(2) building user-adaptive and collaborative AI systems that can align with diverse user goals, preferences, and values in real-world interaction.

This includes personalized and pluralistic alignment, as well as improving models' ability to proactively seek information, infer user goals and intents from interaction, and adapt their responses and actions responsibly.

During Summer 2025, I was a Research Engineer Intern at the Center for AI Safety. Previously, I was a research assistant at Dartmouth College where I had the chance to work with Dr. Ruibo Liu and Prof. Soroush Vosoughi on value alignment for LLMs.

Email  /  Github  /  Google Scholar  /  Linkedin  /  X

A benchmark and set of analyses for evaluating whether vision-language models respect contextual integrity in location disclosure for image geolocation, revealing that violations of contextual norms may result in privacy harms, characterized by over-disclosure of sensitive locations, poor privacy-utility tradeoffs, and misalignment with human privacy expectations.

We propose Collaborative Calibration, a collaborative approach to elicit, calibrate, and rationalize prediction confidence of LLMs.

Alignment training with data from multi-LLM simulated social interactions, as an efficient, effective, and stable alternative for RLHF.

In this work, we present a novel visual analytical framework to support the analysis of transformer-based generative models.

Competed for Shared Task 0: Generalization and Typologically Diverse Morphological Inflection and achieved the highest performance among all submission in both small and large training conditions.

Credits to Jon Barron's website: source code.