CV
PhD Candidate, University of Michigan · LLM Personalization & Pluralistic Alignment
yinuoxu@umich.edu · Ann Arbor, MI · LinkedIn · GitHub · Google Scholar
Education
University of Michigan, Ann Arbor 2023 – Expected May 2028
Ph.D. in Information (GPA: 3.80/4.00) · Advisor: David Jurgens
University of California, Berkeley 2023
B.A. in Data Science & Statistics (GPA: 3.83/4.00)
Awards, Grants & Leadership
- Co-Organizer, Personalization Without Prejudice (PwP) workshop, EMNLP 2027
- Organizer, NLP@Michigan Day 2026, 2025–2026
- Organizer, Michigan NLP Reading Group, 2025–2026
- Science Communication Fellow (selected as 1 of 20 fellows university-wide), 2023–2024
- University of Michigan School of Information Merit Fellowship, 2023
Publications (Selected)
- Under Review (ARR) Xu, Y., Adams, J., Hu, B., Kumar, S., Jurgens, D., & Basharat, A. (2026). VISTA: Steering LLMs with latent user values inferred from a few survey answers.
- EMNLP 2026 · Oral Xu, Y., & Jurgens, D. (2026). Beyond consensus: Perspectivist modeling and evaluation of annotator disagreement in NLP.
- ACL 2026 · Oral Xu, Y., Derricks, V., Earl, A., & Jurgens, D. (2026). Modeling annotator disagreement with demographic-aware experts and synthetic perspectives.
- NAACL 2025 Xu, Y., Chen, H., Rakshit, S., et al., & Jurgens, D. (2025). Causally modeling the linguistic and social factors that predict email response. (Equal contribution)
Relevant Experience
Research & Development Intern, Kitware May 2026 – Aug 2026
Project: VISTA — Steering LLMs with Latent User Values Inferred from a Few Survey Answers
- Led development of VISTA, which infers a probabilistic user-value profile from survey answers (variational encoder) and maps it via a hypernetwork to steering vectors and soft prompts to steer a frozen Llama-3-8B for personalized alignment in choice selection and free-text generation (soft prompts refined with personalized DPO). Matched or beat in-context profile prompting on 3 datasets (up to ∼10 accuracy points) with no added prompt tokens; in free-text generation, using only survey answers, outperformed prompting with answers plus demographics, with the best net gain across minority- and majority-view users.
- Designed adaptive questioning that personalizes from just 6% of a survey, choosing the most informative questions and stopping automatically while keeping 94% of the full-survey accuracy gain (vs. at most 42% for standard adaptive testing).
Lead Researcher, University of Michigan School of Information Aug 2023 – Present
Survey: Perspectivist Modeling of Annotator Disagreement (EMNLP 2026 · Oral)
- Synthesized 110+ works on annotator disagreement in subjective, safety-relevant tasks (e.g., toxicity) into a unified taxonomy of disagreement sources and a framework of prediction targets and pooling structures, identifying the modeling gaps that motivate Project 2.
Project 1: Modeling Annotator Disagreement with Demographic-Aware Experts (ACL 2026 · Oral)
- Designed Demographic-Aware Mixture-of-Experts, an interpretable architecture with demographic-aware routing to capture structured variation in annotation behavior and intersectional perspectives.
- Achieved state-of-the-art subgroup-level performance across five benchmark datasets, improving fairness for underrepresented groups under data imbalance and sparse subgroup data.
- Evaluated LLM annotation reliability through zero-shot, few-shot, and LoRA-finetuned experiments, measuring alignment between persona-prompted LLM ratings and human judgments.
- Developed data-efficient pipelines blending real and LLM-generated synthetic annotations with alignment-weighted loss; implemented multi-GPU training workflows for model scaling and reproducibility.
Project 2: Bayesian Hierarchical Adapter Personas
- Designing a hypernetwork-based framework that generates parameter-efficient LoRA adapters from task, data, and annotator-feature interactions, with a three-level Bayesian hierarchy (population, subpopulation, individual) for pluralistic alignment of model behavior with group and individual norms.
- Leading evaluation on a 140K-instance benchmark spanning normative, pragmatic, and preference-judgment tasks, against Jury Learning, Mixture-of-Personas, and sociodemographic prompting baselines.
Project 3: Measuring Racial Disparities in Doctor-Patient Conversations
- Built a zero-shot LLaMA-70B labeling pipeline for doctor communication quality, validated against human ratings; designed a causal mediation model (doctor language → patient behavior → outcome); in prep for PNAS.
NLP Intern, Evisort (acquired by Workday) May 2021 – Aug 2021
- Trained LightGBM and bi-LSTM NER models (TensorFlow) for contract field extraction, deployed to production; built sensitive-data redaction pipelines with the Google DLP API.
Teaching Experience
- Becoming a Data Scientist (SI 568), Graduate Student Instructor, University of Michigan, Winter 2026
- Applied Machine Learning (SI 670), Graduate Student Instructor, University of Michigan, Fall 2025
- Natural Language Processing (Info 159/259), Teaching Assistant, UC Berkeley, Spring 2023
- Principles and Techniques of Data Science (Data 100/200), Undergraduate Student Instructor, UC Berkeley, Fall 2022
- Principles and Techniques of Data Science (Data 100/200), Undergraduate Student Instructor, UC Berkeley, Summer 2022
- Technology and Social Impact (EECS 198), Lead Facilitator, UC Berkeley, Spring 2023
Professional Service
- Program Committee Member, International Conference on Computational Social Science (IC2S2), 2026
- Program Committee Member, Workshop on NLP and Computational Social Science (at ACL), 2026
Technical Skills
Python · PyTorch · Hugging Face Transformers · PEFT / LoRA · vLLM · SLURM · Git · Hypernetworks · Activation Steering · DPO · Mixture-of-Experts · Multi-GPU Training (DDP) · Bayesian & Variational Inference · Causal Inference · LLM Evaluation & Alignment · Optuna
