Hanqi Xiao

I am a undergraduate research assistant at MURGe-Lab advised by Prof. Mohit Bansal, and Prof. Elias Stengel-Eskin. I am an undergraduate at UNC Chapel Hill. I'm applying this cycle for fall 2026 PhD programs.

I am interested in exploring long term, adaptable, and updatable memories in LLMs that remain efficient through software hardware codesign, moving a bit closer to a machine that can experience remember and pursue long term goals while ensuring robustness and safety. I believe building in a sense of history and continuity into LLMs will solve subtle capability gaps less visible through current evals.

Recently news: Honorable Mention for the CRA Outstanding Undergraduate Researcher Award (Dec 2025).

Email / LinkedIn / Twitter / Github / CV / Research Plan

Research

I'm interested in understanding generalizable principles/insights to improve compression, robustness/interpretability, and embedded applications. I have recently been working on LLM quantization and predictability to promote efficiency and robustness.

	Generalized Correctness Models: Learning Calibrated and Model-Agnostic Predictors from Historical Patterns Hanqi Xiao, Vaidehi Patil, Hyunji Lee, Elias Stengel-Eskin, Mohit Bansal arXiv preprint, 2025 Link We propose Generalized Correctness Models (GCMs) trained on multiple models’ historical correctness data, to yield calibrated, model-agnostic predictors of answer reliability. Unlike prior self-knowledge approaches, GCMs leverage patterns from multiple models and outperform individual models’ confidence estimates in selective prediction and calibration tasks.
	Task-Circuit Quantization: Leveraging Knowledge Localization and Interpretability for Compression Hanqi Xiao, Yi-Lin Sung, Elias Stengel-Eskin, Mohit Bansal, COLM, 2025, top 3.35% of accepted papers by review (7,7,9). Link Using interpretability informed saliency scores based on task-specific information to localize important weights to preserve during model compression, yielding improvements for both general and task specific quantization
	Developing a Prototype for the Systems Literature Analysis Engine (SLAE) Hanqi Xiao, Robert Peters, Davyd Voloshyn, Jonathan Vester, Thomas Bland, Kristen Hassmiller-Lich 46th Annual Meeting of the Society for Medical Decision Making, 2024 Link Poster presentation on developing a prototype for the Systems Literature Analysis Engine, focusing on automating the connection of scientific evidence to inform causal loop diagrams on 'wicked' problems and systems.
	Using systems thinking modalities to map the mental health crisis at UNC: a national case study Aaron Carpenter, Hanqi Xiao, Lily Goldberg, Rachel Smith, Riley Harper, Deborah Shoola APHA 2023 Annual Meeting and Expo, 2023 Link We investigate the causal elements that influence mental health trends among university students using systems thinking. Our research includes a literature review of 1000+ articles (aided by a custom programatic search and categorization system we develop), student surveys (n=77), and interviews with students, staff, and faculty to identify intervention models and leverage points for combating the suicide crisis on campus. Map the System 2023 global finalist project (Skoll Centre, Oxford University).

Community

I previously founded and led AI@UNC, the undergraduate student organization for AI at UNC-Chapel Hill. We hosted 150+ members in good standing and supported 14 completed projects. I also co-chair the mainly advisory board of SAILea, an organization supporting 70(as of Oct 2025) highschool AI clubs. I am interested in Artificial Human Intelligence, and like to talk to other people who are interested as well.

Gallery

Website source borrowed from here.