Publications

Strategy-Aware Optimization Modeling with Reasoning LLMs

Published in ICML, 2026

This paper introduces SAGE, a strategy-aware framework that makes modeling strategies explicit via solver-verified multi-strategy data and Segment-Weighted GRPO. It significantly boosts the reliability, correctness, and solver efficiency of LLMs in automated optimization modeling, outperforming strong baselines across eight benchmarks.

Recommended citation: @article{zhao2026strategy, title={Strategy-Aware Optimization Modeling with Reasoning LLMs}, author={Zhao, Ruiqing and Li, Fengzhi and Zuo, Yuan and Liu, Rui and Liu, Yansong and Ma, Yunfei and Meng, Fanyu and Feng, Junlan}, journal={arXiv preprint arXiv:2605.02545}, year={2026} } https://arxiv.org/abs/2605.02545

Beyond One-Size-Fits-All: Adaptive Subgraph Denoising for Zero-Shot Graph Learning with Large Language Models

Published in SIGKDD, 2026

To address the noise and suboptimality of fixed subgraph extraction in zero-shot graph reasoning, we propose GraphSSR, a framework that utilizes a “Sample-Select-Reason” (SSR) pipeline, supervised fine-tuning (SSR-SFT), and reinforcement learning (SSR-RL) to enable Large Language Models to adaptively extract and denoise task-relevant graph structures for superior generalization.

Recommended citation: @article{li2026beyond, title={Beyond One-Size-Fits-All: Adaptive Subgraph Denoising for Zero-Shot Graph Learning with Large Language Models}, author={Li, Fengzhi and Zhang, Liang and Zuo, Yuan and Zhao, Ruiqing and Liu, YanSong and Ma, Yunfei and Meng, Fanyu and Feng, Junlan}, journal={arXiv preprint arXiv:2603.02938}, year={2026} } https://arxiv.org/abs/2603.02938

Memory-Tuning: A Unified Parameter-Efficient Tuning Method for Pre-Trained Language Models

Published in IEEE Transactions on Audio, Speech and Language Processing, 2024

We propose memory-tuning, a novel parameter-efficient method that unifies task-specific knowledge learning for both multi-head attention and feed-forward networks in Transformers, theoretically linking it to prefix tuning while outperforming full fine-tuning on eight benchmarks across sentence- and token-level tasks.

Recommended citation: @article{qi2024memory, title={Memory-Tuning: A Unified Parameter-Efficient Tuning Method for Pre-Trained Language Models}, author={Qi, Wang and Liu, Rui and Zuo, Yuan and Li, Fengzhi and Chen, Yong and Wu, Junjie}, journal={IEEE Transactions on Audio, Speech and Language Processing}, volume={33}, pages={1--10}, year={2024}, publisher={IEEE} } https://ieeexplore.ieee.org/abstract/document/10769026

LLMs as Zero-shot Graph Learners: Alignment of GNN Representations with LLM Token Embeddings

Published in 38th Conference on Neural Information Processing Systems (NeurIPS 2024), 2024

We propose TEA-GLM, a zero-shot graph learning framework that aligns GNN representations with LLM token embeddings via a fixed linear projector, enabling cross-task and cross-dataset generalization without LLM fine-tuning.

Recommended citation: @article{wang2024llms, title={Llms as zero-shot graph learners: Alignment of gnn representations with llm token embeddings}, author={Wang, Duo and Zuo, Yuan and Li, Fengzhi and Wu, Junjie}, journal={Advances in neural information processing systems}, volume={37}, pages={5950--5973}, year={2024} } https://proceedings.neurips.cc/paper_files/paper/2024/file/0b77d3a82b59e9d9899370b378087faf-Paper-Conference.pdf

BoostXML: Gradient Boosting for Extreme Multilabel Text Classification With Tail Labels

Published in IEEE Transactions on Neural Networks and Learning Systems, 2023

We present BoostXML, a deep learning-based extreme multilabel text classification method enhanced by gradient boosting, which specifically improves tail-label prediction through a Boosting Step that optimizes residuals from unfitted tail-label instances, a Corrective Step to avoid optimization mismatches, and a Pretraining Step to balance label focus.

Recommended citation: @article{li2023boostxml, title={Boostxml: gradient boosting for extreme multilabel text classification with tail labels}, author={Li, Fengzhi and Zuo, Yuan and Lin, Hao and Wu, Junjie}, journal={IEEE Transactions on Neural Networks and Learning Systems}, volume={35}, number={11}, pages={15292--15305}, year={2023}, publisher={IEEE} } https://ieeexplore.ieee.org/abstract/document/10161991

Where to go? Predicting next location in IoT environment

Published in Frontiers of Computer Science, 2020

We propose TSIS, a neural model that jointly learns from trajectory transitions and IoT signal sequences via gated GNNs and GRUs for accurate next-location prediction in IoT environments.

Recommended citation: @article{lin2021go, title={Where to go? Predicting next location in IoT environment}, author={Lin, Hao and Liu, Guannan and Li, Fengzhi and Zuo, Yuan}, journal={Frontiers of Computer Science}, volume={15}, number={1}, pages={151306}, year={2021}, publisher={Springer} } https://link.springer.com/article/10.1007/s11704-019-9118-9

Matching Neural Network for Extreme Multi-Label Learning

Published in 4th International Conference on Artificial Intelligence Applications and Technologies (AIAAT 2020), 2020

We propose MNN, a neural matching framework for extreme multi-label learning that maps features and labels into aligned representations via contrastive learning, overcoming tail-label challenges.

Recommended citation: @inproceedings{zhao2020matching, title={Matching Neural Network for Extreme Multi-Label Learning}, author={Zhao, Zhiyun and Li, Fengzhi and Zuo, Yuan and Wu, Junjie}, booktitle={Journal of Physics: Conference Series}, volume={1642}, number={1}, pages={012013}, year={2020}, organization={IOP Publishing} } https://iopscience.iop.org/article/10.1088/1742-6596/1642/1/012013/meta

Fengzhi Li