I'm Hua (Edward) XU (徐画), a final-year Data Science undergraduate at HKUST(GZ), graduating in June 2027. My interests lie in computational neuroscience, NeuroAI, and probabilistic machine learning.
I am interested in how learning shapes computation in brains and artificial systems. When two systems behave similarly, what evidence would show that they also compute in similar ways? I hope to study this through models of neural dynamics, comparisons between brains and AI, and the limits of what we can infer from observations and interventions.
My current work spans controllable discrete diffusion with probabilistic circuits (the COFFEE paper is now an arXiv preprint) and exploratory comparisons between diffusion language models and fMRI responses. Earlier projects include time-series modeling with Prof. Yutao Yue, image generation with Prof. Andrew Luo, and synthetic biology with Prof. Julie Qiaojin LIN.
I lead COFFEE, a framework for probabilistic-circuit guidance in constrained discrete generation, studying how explicit sequence objectives can be incorporated into sampling and where exact inference gives way to approximation.
In collaboration with Gwen Yidou Weng and Prof. Anji Liu and Prof. Yuwen Huang and Prof. Guy Van den Broeck
I compare diffusion-language-model representations with human fMRI responses, using held-out evaluation and controlled comparisons to examine what predictive alignment can tell us about shared computation.
Supervised by Prof. Julie Qiaojin Lin and Prof. Wei Wang and Prof. Kaishun Wu
I process multi-electrode-array (MEA) recordings, with attention to data quality, provenance, and longitudinal neural dynamics across physical wells.
Supervised by Prof. Julie Qiaojin Lin and Prof. Wei Wang and Prof. Kaishun Wu
I am developing a thesis direction on how generative sequence models relate to replay, memory, and the simulation of future trajectories.
Supervised by Prof. Wei Wang and Prof. Kefei Liu
ICLR 2027 submission 2026 arXiv preprint
Summary COFFEE uses a target-free carrier and a compiled finite-state objective model to guide discrete diffusion with sequence-level preferences without retraining the backbone.
My role First author and project lead; formulated the method, implemented the framework, designed evaluations across symbolic, language, and biological tasks, and led the manuscript.
43rd International Conference on Machine Learning (ICML) 2026 Accepted
Summary This work constructs explainable rules for scientific discovery using LLMs. Specifically, we develop a unified framework integrating LLMs with probabilistic modeling through four stages: rule generation via LLM, weight learning through logistic regression, iterative refinement, and evaluation, achieving higher rule quality and better rule combination effects.
My role Equal-contribution core contributor; worked on probabilistic integration for combining LLM-generated rules, evaluation design, and paper writing.
AAAI 2026 Undergraduate Consortium 2026 Accepted proposal
Summary Single-author Undergraduate Consortium research proposal on developing interpretable and trustworthy neural signal decoding systems.
My role Single-author undergraduate consortium proposal; formulated the problem framing and research agenda for interpretable, trustworthy neural signal decoding.
42nd International Conference on Machine Learning (ICML) 2025 Accepted
Summary Leveraging visual pretrained masked autoencoders to address irregular multivariate time series prediction challenges by converting sparse data into time × channel image-like patches, capturing cross-channel interactions with superior accuracy and strong few-shot performance.
My role Equal-contribution core contributor; helped design and run experiments, write the paper, and participate in the rebuttal process.
