Haolun Wu 吳昊倫

I am a Ph.D. candidate in Computer Science at McGill University and Mila - Quebec AI Institute. I am delighted in working with Steve Liu and Fernando Diaz. I also collaborate with Laurent Charlin and Joelle Pineau and have them on my supervisory committee. In 2025, I was a visiting scholar at Stanford Trustworthy AI Research (STAIR) Lab led by Sanmi Koyejo working on trustworthy foundation models, compound AI system alignment, LLMs for personalized adaption, and LLMs for education.

My research centers on learning from human feedback using ML techniques to make AI systems trustworthy, responsible, and align with human needs. My work on human-AI alignment spans both micro-level aspects (such as personalization and data values) and macro-level aspects (such as social goods and norms). Additionally, I love interdisciplinary research and am particularly interested in applying AI/ML techniques to Education and Psychology. Additionally, I am one of the organizers of the OracleLLM community, where we dive into the frontier of using LLMs as oracles—trusted agents capable of providing reliable, high-level insights. I am also a member of the organizing committee for NICE (NLP Academic Exchange Platform), a large community for NLP researchers to share their work and ideas. We organize events in both Mandarin and English. Feel free to reach out and fill this form if you would like to present your research or engage with our community.

During my Ph.D. journey, I interned and collaborated closely with researchers at Microsoft Research, Samsung AI Research, Google Research, and Google DeepMind. My research is generously supported by McGill Graduate Excellence Award, Microsoft Research & Mila Research Grant (twice), MITACS Accelerate Fellowship, Borealis AI Fellowship, Samsung Global Research Outreach Grant, and FRQNT PhD Scholarship (ranked 1st place).

📑 Curriculum Vitae

Past and present research topics include but not limited to:

Machine Learning from Human Feedback. I focus on the fundamental challenge of aligning AI with complex human intent by developing probabilistic and adaptive frameworks that learn from implicit, noisy, or ambiguous feedback. I pioneered DUR to model user preferences as distributions rather than static points, and IMCAT to disentangle multi-faceted intents via contrastive learning. To ensure robust alignment, I frame LLM adaptation as a label noise correction problem in Plugin and developed the TIL mechanism to reweight noisy training samples based on informativeness. Additionally, I design metrics like GA-SS to quantify search success and satisfaction across diverse user groups.

Trustworthy AI and Responsible Systems. I address the critical need for AI systems that are fair, transparent, and safe at a societal scale. Moving beyond isolated models, I view alignment as a multi-stakeholder challenge involving users, creators, and platforms. I developed multi-objective optimization frameworks such as Multi-FR to achieve Pareto-efficient fairness and JMEFairness to mitigate exposure bias. To enhance trust, I focus on scrutability, creating transparent, natural-language user representations via TEARS. Recently, I worked on System-level DPO to align compound AI systems, ensuring safety properties emerge from component interactions.

Knowledge Extraction and Reasoning. My research advances the reliability and reasoning capabilities of Foundation Models by grounding them in structured knowledge. I investigate methods like MuSEE, enabling LLMs to function as reliable knowledge extraction and knowledge base construction tools. I also explore the intersection of graphs and learning, integrating knowledge graphs via hyperbolic embeddings in Hyper-Know, and capturing sequential dynamics through Diffusion-based Contrastive Learning. I further define the frontiers of efficient and adaptive inference through comprehensive research on Geometric Knowledge Distillation, Retrieval-Augmented Generation, and Test-Time Scaling.

Human-centered AI in Education and Collaboration. I build intelligent systems that scaffold learning and collaboration in high-stakes environments. I developed SSRLBot, an LLM-based agent that uses socially shared regulation to guide medical team diagnostics. I ground these systems in multimodal analysis, examining how facial expressions reveal distinct novice-expert patterns and utilizing physiological signals to identify critical moments in team decision-making. Additionally, I show how social annotations foster psychological safety in collaborative reading. Beyond education, I investigate Multi-Agent Debate and hierarchical frameworks like PartnerMAS to enable cooperative and transparent reasoning.