I am currently a second-year Ph.D. student at the School of Information and Communication Engineering, University of Electronic Science and Technology of China(UESTC, 电子科技大学) supervised by Prof. Fanman Meng (孟凡满).
My research primarily centers on continual learning and the efficient reuse of multimodal large language models and vision-language models. Earlier in my Ph.D. studies, I focused on class-incremental learning. More recently, I have become interested in model merging, exploring how the efficient integration of pre-trained models can enhance multi-task performance and support continual model adaptation.
🔥 News
- 2025.05: 🎉 One co-author paper has been accepted to ICIP 2025.
- 2025.01: 🎉 One co-author paper have been accepted to TMM 2025.
- 2024.12: 🎉 One co-author paper has been accepted to ICASSP 2025.
- 2024.06: 🥳 I went to Seattle, USA to attend the CVPR 2024 conference.
- 2024.02: 🎉 One first-author paper has been accepted to CVPR 2024.
- 2023.12: 🎉 One co-author paper has been accepted to IEEE Sensors Journal 2024.
- 2023.05: 🎉 One co-author paper has been accepted to Neurocomputing 2023.
- 2023.03: 🎉 One co-author paper has been accepted to ICASSP 2023.
- 2022.12: 🎉 One first-author paper has been accepted to Neurocomputing 2023.
- 2022.06: 🎉 One co-author paper has been accepted to ICIP 2022.
- 2021.10: 🎉 My first paper has been accepted to SPIE Medical Imaging 2022.
📝 Publications
Dual-consistency Model Inversion for Non-exemplar Class Incremental Learning
Zihuan Qiu, Yi Xu, Fanman Meng, Hongliang Li, Linfeng Xu, Qingbo Wu
- DCMI: Uses semantic + domain consistency to generate high-fidelity synthetic old-class images, sharply cutting domain shift.
- Prototypical Routing: Leverages class prototypes to reduce bias and lift old-class accuracy, achieving state-of-the-art results.
ISM-Net: Mining incremental semantics for class incremental learning
Zihuan Qiu, Linfeng Xu, Zhichuan Wang, Qingbo Wu, Fanman Meng, Hongliang Li
- Incremental Semantics Mining (ISM) – strips old-class semantics from new-class features, sharply cutting old–new confusion and catastrophic forgetting.
- Lightweight Expansion + Old-Model Queue – distillation adds a compact extra feature space while a rolling queue preserves early knowledge, delivering SOTA accuracy with minimal memory/compute on CIFAR-100 and ImageNet.
TMM 2025[Distribution-Level Memory Recall for Continual Learning: Preserving Knowledge and Avoiding Confusion], Shaoxu Cheng, Kanglei Geng, Chiyuan He, Zihuan Qiu, Linfeng Xu, Heqian Qiu, Lanxiao Wang, Qingbo Wu, Fanman Meng, Hongliang Li.ICASSP 2025[GFR: Generic feature representations for class incremental learning], Yukun Chen, Zihuan Qiu, Fanman Meng, Hongliang Li, Linfeng Xu, Qingbo Wu.IEEE Sensors Journal 2024[Continual Egocentric Activity Recognition with Foreseeable-Generalized Visual-IMU Representations], Chiyuan He, Shaoxu Cheng, Zihuan Qiu, Linfeng Xu, Fanman Meng, Qingbo Wu, Hongliang Li.Neurocomputing 2023[GFR: Generic feature representations for class incremental learning], Zhichuan Wang, Linfeng Xu, Zihuan Qiu, Qingbo Wu, Fanman Meng, Hongliang Li.ICASSP 2023[MFAT: A multi-level feature aggregated transformer for person re-identification],Bowen Tan, Linfeng Xu, Zihuan Qiu, Qingbo Wu, Fanman Meng.ICIP 2022[Eldnet: Establishment and refinement of edge likelihood distributions for camouflaged object detection], Chiyuan He, Linfeng Xu, Zihuan Qiu.SPIE Medical Imaging 2022BDG-Net: boundary distribution guided network for accurate polyp segmentation, Zihuan Qiu, Zhichuan Wang, Miaomiao Zhang, Ziyong Xu, Jie Fan, Linfeng Xu.
Educations
- 2023.09 - present, Phd student, Information and Communication Engineering, University of Electronic Science and Technology of China(UESTC), Chengdu.
- 2020.09 - 2023.06, Master student, Electronic Information, University of Electronic Science and Technology of China(UESTC), Chengdu.
- 2015.09 - 2019.06, Undergraduate, Communication Engineering, Tiangong University, Tianjing.
Services
Reviewers
- NeurIPS 2025, MM 2025, ICME 2025, ICLR 2025, ICASSP 2025