Zixuan (sounds like “Six-one”) Xia

I am a PhD student in Computer Science at the University of Neuchâtel, supervised by Prof. Lydia Y. Chen. I received my Master’s degree in Computer Science from the University of Bern in September 2026, and my Bachelor’s degree in Software Engineering from Xi’an Jiaotong University in China. My research interests lie in deep learning optimization and training dynamics, trustworthy generative AI, and multimodal representation learning.


Publications

Understanding Confabulation and Rethinking Reconstruction in Activation Explanations

Understanding Confabulation and Rethinking Reconstruction in Activation Explanations

Gert Lek, Zixuan Xia, Pin-Yu Chen, Lydia Y. Chen
arXiv, 2026

Flow-NLA moves beyond point reconstruction in Natural Language Autoencoders by modeling distributions of activations compatible with an explanation, retaining utility gains while reducing confabulation and writing defects.

Exact Regret Frontiers and Externality Scheduling in Centralized Serial-Dictatorship Bandits

Exact Regret Frontiers and Externality Scheduling in Centralized Serial-Dictatorship Bandits

Lishang Xu, Guodong Ma, Pengcheng Weng, Zixuan Xia
arXiv, 2026

A theoretical study of centralized serial-dictatorship matching bandits, characterizing exact regret frontiers and showing how exploration scheduling externalities affect attainable regret.

COMPASS: Complete Multimodal Fusion via Proxy Tokens and Shared Spaces for Ubiquitous Sensing

COMPASS: Complete Multimodal Fusion via Proxy Tokens and Shared Spaces for Ubiquitous Sensing

Hao Wang*, Yanyu Qian*, Pengcheng Weng, Zixuan Xia, William Dan, Yangxin Xu, Fei Wang
arXiv, 2026

A missing-modality multimodal fusion framework that preserves a modality-complete fusion interface by synthesizing proxy tokens for absent modalities in a shared latent space.

When Gradient Optimization Is Not Enough: † Dispersive and Anchoring Geometric Regularizer for Multimodal Learning

When Gradient Optimization Is Not Enough: † Dispersive and Anchoring Geometric Regularizer for Multimodal Learning

Zixuan Xia*, Hao Wang*, Pengcheng Weng*, Yanyu Qian, Yangxin Xu, William Dan, Fei Wang
TMLR, 2026

A plug-and-play geometry-aware regularization framework for multimodal learning, encouraging intra-modal dispersion and inter-modal anchoring to mitigate representation collapse and cross-modal drift.

K-Score: Kalman Filter as a Principled Alternative to Reward Normalization in Reinforcement Learning

K-Score: Kalman Filter as a Principled Alternative to Reward Normalization in Reinforcement Learning

Zixuan Xia*, Quanxi Li*
NewInML Workshop, ICML 2025

A lightweight Kalman-filter-based reward estimation method for policy gradient reinforcement learning, replacing fixed reward normalization with online latent reward mean estimation to reduce variance and improve convergence.

KOALA++: Efficient Kalman-Based Optimization with Gradient-Covariance Products

KOALA++: Efficient Kalman-Based Optimization with Gradient-Covariance Products

Zixuan Xia, Aram Davtyan, Paolo Favaro
NeurIPS, 2025

An extension of KOALA, a neural network optimization algorithm based on Kalman filtering, with implicit full weights covariance matrix.