Zhiwei Li

Zhiwei Li (李志伟) is a PhD candidate at the Australian Artificial Intelligence Institute (AAII), University of Technology Sydney, advised by Prof. Guodong Long. He studies how individual users can retain personalization, evidence, and control of their own data in the foundation-model era. His earlier work on Federated Recommendation pursued privacy-preserving personalization by separating shared and locally personalized representations. He is now building On-device Personal Agents — user-centric, privacy-controlled, interpretable, and continually-learning, that serve as reliable mediators between users and external knowledge, content, and recommender systems. His work appears at ICLR, AAAI, IJCAI, and IEEE TKDE.

Previously, he received his M.S. in Computer Science from ShanghaiTech University in 2023, advised by Prof. Lu Sun, and earned his B.E. from Zhengzhou University in 2020 as a member of the ZZU-DROID bipedal-robot team.

Research Interests

2026 – present: User-Centric Personal Intelligent Systems — on-device personal agentic AI under user control
2023 – present: Federated Recommendation Systems — privacy-preserving personalization
2020 – 2023: Multi-View / Multi-Label Learning — learning under partial labels and incomplete views

On the Academic Job Market — 2026/27

Zhiwei is graduating from UTS AAII in mid-2027 and is actively seeking Postdoctoral Research Fellow and Tenure-Track / Faculty positions across Australia, Singapore, Hong Kong, and mainland China, with a focus on On-device Personal Agentic AI, Federated Recommendation, Privacy-Preserving Foundation models, and User-centric Intelligent Systems. Conversations and collaborations are very welcome — please feel free to reach out.

News

May 01, 2026	Our survey paper on Personalized Federated Foundation Models for Recommendation is accepted by IJCAI–ECAI 2026 (Survey Track)!
Nov 15, 2025	We have two papers accepted by AAAI 2026!
Apr 15, 2025	I am honored to be selected as one of the Best Reviewers for AISTATS 2025!
Dec 15, 2024	We have a tutorial on Personalized Federated Recommendation Systems accepted by WWW 2025.
Dec 10, 2024	We have one paper accepted by AAAI 2025!

Selected Publications

IJCAI 2026
A Survey of Personalized Federated Foundation Models for Privacy-Preserving Recommendation

Zhiwei Li, Guodong Long, Chunxu Zhang, and 3 more authors

In Proceedings of the 35th International Joint Conference on Artificial Intelligence (Survey Track), 2026

Abs arXiv Bib 1

This survey reframes the privacy challenge in federated recommendation from an architectural problem (data stays local) to a geometric one: gradient updates must be orthogonal to a sensitive subspace in the latent representation space. Building on this perspective, we organize the rapidly growing literature on Personalized Federated Foundation Models for recommendation into a unified framework, identify five open problems (semantic gap, structural rigidity, semantic heterogeneity, optimization conflict, compute and communication), and lay out five future directions that map one-to-one to those problems.
@inproceedings{li2026pffm, title = {A Survey of Personalized Federated Foundation Models for Privacy-Preserving Recommendation}, author = {Li, Zhiwei and Long, Guodong and Zhang, Chunxu and Zhang, Honglei and Zhang, Chengqi and Jiang, Jing}, booktitle = {Proceedings of the 35th International Joint Conference on Artificial Intelligence (Survey Track)}, year = {2026}, }
AAAI 2026
Federated Vision-Language-Recommendation with Personalized Fusion

Zhiwei Li, Guodong Long, Jing Jiang, and 2 more authors

In Proceedings of the AAAI Conference on Artificial Intelligence, 2026

Abs arXiv Bib Code 10

FedVLR is a federated recommendation framework that integrates vision-language models with personalized fusion to address multimodal cold-start and heterogeneity. It uses adapter tuning on a frozen VLM backbone and a per-client fusion head that is trained locally, retaining global semantic priors while adapting to local taste distributions.
@inproceedings{li2026fedvlr, title = {Federated Vision-Language-Recommendation with Personalized Fusion}, author = {Li, Zhiwei and Long, Guodong and Jiang, Jing and Zhang, Chengqi and Yang, Qiang}, booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence}, year = {2026}, }
AAAI 2025
Personalized Federated Collaborative Filtering: A Variational AutoEncoder Approach

Zhiwei Li, Guodong Long, Tianyi Zhou, and 2 more authors

In Proceedings of the AAAI Conference on Artificial Intelligence, 2025

Abs arXiv Bib Code 31

FedDAE casts personalized federated collaborative filtering as a variational autoencoder problem in which each client has its own decoder while sharing an encoder. This decomposition aligns the personalization-vs-generalization trade-off with the latent-space structure of the VAE and yields consistent gains across MovieLens, LastFM, and HetRec benchmarks.
@inproceedings{li2025feddae, title = {Personalized Federated Collaborative Filtering: A Variational AutoEncoder Approach}, author = {Li, Zhiwei and Long, Guodong and Zhou, Tianyi and Jiang, Jing and Zhang, Chengqi}, booktitle = {Proceedings of the AAAI Conference on Artificial Intelligence}, year = {2025}, }
ICLR 2024
Federated Recommendation with Additive Personalization

Zhiwei Li, Guodong Long, and Tianyi Zhou

In International Conference on Learning Representations, 2024

Abs arXiv Bib HTML Code 88

FedRAP decomposes a recommendation model into a globally shared component and a per-client additive personalization term, regularized to be sparse so that personalization is communicated efficiently. The method delivers state-of-the-art performance on standard federated recommendation benchmarks while using a fraction of the per-round communication.
@inproceedings{li2024fedrap, title = {Federated Recommendation with Additive Personalization}, author = {Li, Zhiwei and Long, Guodong and Zhou, Tianyi}, booktitle = {International Conference on Learning Representations}, year = {2024}, }