Scholar
Zexi Li
Google Scholar ID: 6lMg5eoAAAAJ
Alibaba Group
Citations & Impact
Citations
904
H-index
14
i10-index
18
Publications
20
Co-authors
4
list available
Contact
Publications
Resume (English only)
Academic Achievements
- Paper “Improving Model Fusion by Training-time Neuron Alignment with Fixed Neuron Anchors” accepted by IEEE TPAMI (Oct 2025)
- Paper “WISE: Rethinking the Knowledge Memory for Lifelong Model Editing of Large Language Models” accepted at NeurIPS 2024
- Paper “Editing as Unlearning: Are Knowledge Editing Methods Strong Baselines for Large Language Model Unlearning?” accepted at NeurIPS 2025 Workshops (Lock-LLM & LLM Evals)
- Paper “FedGuCci: Making Local Models More Connected in Landscape for Federated Learning” accepted at KDD 2025
- Paper “Revisiting Weighted Aggregation in Federated Learning with Neural Networks” accepted at ICML 2023
- Paper “Can We Share Models If Sharing Data Is Not an Option?” published in Patterns (Cell Press)
- Paper “No Fear of Classifier Biases: Neural Collapse Inspired Federated Learning with Synthetic and Fixed Classifier” accepted at ICCV 2023
- Paper “Towards Effective Clustered Federated Learning: A Peer-to-peer Framework with Adaptive Neighbor Matching” published in IEEE Transactions on Big Data
- Paper “Resource-Efficient Knowledge Editing for Mobile LLMs” won Best Poster Award at MobiUK 2025
- Invited as Session Chair for KDD 2025
Background
- Research Scientist at Tongyi Lab, Alibaba Group
- Main research interests include: Large Language Models (LLMs) and agentic intelligence (LLM agents, reasoning, multi-agent systems)
- Model editing and memory management for LLMs
- Parametric understanding of LLMs (localization, merging, scaling, pruning, stitching, unlearning, editing, etc.)
- Text-to-model generation
- Trustworthy deep learning: privacy-preserving federated learning, efficient & robust algorithm design, generalization, personalization, and training dynamics
- Mechanistic interpretability of neural networks (weight decay, loss landscape, permutation invariance, linear mode connectivity, etc.)
- Socio-technical issues in collaborative learning
- Responsible and trustworthy AI
