arXiv preprint
Zehao Xiao
Postdoctoral Researcher, Huawei Noah's Ark Lab, Paris
Ph.D., University of Amsterdam (VIS Lab & AIM Lab)
About
I'm currently a postdoctoral researcher at Huawei Noah's Ark Lab in Paris, where I work on pre-training and post-training of time series foundation models. Before that, I completed my Ph.D. in the VIS Lab & AIM Lab at the University of Amsterdam, supervised by Cees Snoek and co-supervised by Xiantong Zhen. During my Ph.D. I mainly worked on generalization and test-time adaptation, and I was a member of the ELLIS society. Before joining UvA in 2020, I received my Master's and Bachelor's degrees from Beihang University.
My research centers on one question: how should pretrained models keep adapting after pretraining — at test time, under distribution shift, without labels or retraining? This through-line runs from domain generalization and test-time adaptation, through meta-learning and probabilistic, energy-based methods, to prompt-based adaptation of vision-language models and multimodal alignment. Most recently, I apply it to pre-training and post-training of time series foundation models (TSFM), steering frozen forecasters with in-context learning and test-time optimization.
News
- [Sep 2026] Released PaCTS, instance-adaptive prompts for frozen time-series foundation models, as an arXiv preprint.
- [Sep 2026] Released Tabby, an open pretraining recipe and 145M-parameter time series foundation model, as a technical report (arXiv:2609.13956).
- [Jun 2026] Oral paper accepted to the ICML 2026 Workshop on Foundation Models for Structured Data.
- [May 2026] Papers accepted to ICML 2026, Nature Communications, and TMLR.
- [Jan 2026] One paper accepted to ICLR 2026.
- [Jun 2025] Successfully defended my thesis Learning to Generalize at Test Time.
- [May 2025] One paper accepted to ICML 2025.
- [Jan 2025] One paper accepted to ICLR 2025.
- [Nov 2024] Survey Beyond Model Adaptation at Test Time released; one paper to WACV 2025.
- [Apr 2024] Papers accepted to NeurIPS 2024 and CoLLAs 2024.
- [Feb 2024] One paper accepted to CVPR 2024.
- [2021-2023] Papers at ICLR 2022/2023, NeurIPS 2022/2023, ICML 2021.
Publications
Selected publications. For the full list, see my Google Scholar profile.
2026
Preprint (Technical Report)
ACM Transactions on Knowledge Discovery from Data
Nature Communications
ICML
TMLR
ICLR
2025
ICML
WACV
ICLR
2024
Preprint
NeurIPS
CoLLAs
Probabilistic Test-Time Generalization by Variational Neighbor-Labeling
Variational neighbor labels with meta-generalization: pseudo labels as latent variables, meta-learned with variational inference and neighboring information for test-time domain generalization.
CVPR
Any-Shift Prompting for Generalization over Distributions
A general probabilistic inference framework that uses both training and test task information for prompt generation, improving generalization across distribution shifts.
2023
NeurIPS
ProtoDiff: Learning to Learn Prototypical Networks by Task-Guided Diffusion
A task-guided diffusion model during meta-training that gradually generates prototypes for efficient class representations.
ICLR
Energy-Based Test Sample Adaptation for Domain Generalization
A discriminative energy-based model that adapts target samples to the source domain distributions for domain generalization.
2022
NeurIPS
Association Graph Learning for Multi-Task Classification with Category Shifts
A new multi-task learning setting that addresses category shifts from training to test data.
ICLR
Learning to Generalize across Domains on Single Test Samples
A meta-learning paradigm that learns to adapt with single samples at training time, then further adapts to each single test sample at test time.
2021
ICML
A Bit More Bayesian: Domain-Invariant Learning with Uncertainty
Introduces weight uncertainty via variational Bayesian inference to better explore domain-invariant learning.
2019
ICCV
Relational Attention Network for Crowd Counting
A self-attention mechanism capturing pixel interdependence for crowd counting and density estimation.
CVPR
Crowd Counting and Density Estimation by Trellis Encoder-Decoder Networks
A trellis encoder-decoder network for crowd counting that generates high-quality density maps.
