Research & Projects

Publications

Preprints

Research & Project Experience

My research connects reliable learning and model evaluation with human–AI interaction and applied machine learning.

Affiliation logos: rightmost = my home institution; any to the left = primary collaborator(s).

Learning Systems and Human–AI Interaction

Federated and Heterogeneous Learning

• Data heterogeneity survey — Expert Systems (2026). Led the survey framework on federated learning under non-IID and heterogeneous data, covering healthcare, finance, and IoT case studies; also designed frequency-adaptive semantic distillation with semantic topology and dynamic frequency-based weighting. Advisors: Prof. Qilei Li and Prof. David Camacho.
• FedSAP — arXiv preprint (2026). Structured adaptive partitioning for federated learning on multi-domain heterogeneous edge devices. arXiv:2610.01638. Advisor: Prof. Qilei Li.
• MOSAIC — arXiv preprint (2026). Margin-oriented semantic–appearance interaction correction beyond domain-level adaptation for personalized federated vision–language models. arXiv:2610.01625. Advisor: Prof. Qilei Li.
• After Cooperation Is Learned — arXiv preprint (2026). Gradient routing and optimizer-dependent maintenance of learned cooperation in multi-agent reinforcement learning. arXiv:2610.01630. Advisor: Prof. Qilei Li.
• FedSocratic — submitted to ICLR 2027. Verifiable questioning for federated language models. Advisors: Prof. Qilei Li and Prof. David Camacho.

Unlearning, Context Effects, and Model Evaluation

• Reasoning after Unlearning — first author; submitted to ICLR 2027. Evaluates original, unlearned, and target-excluded retrained Llama checkpoints on TOFU, separating context effects, answer exposure, and use of supplied answer information through matched controls, answer masking, and identity substitution. Advisor: Prof. Qilei Li.
• Multimodal unlearning — co-authored; submitted to ICLR 2027. Auxiliary supervision for unlearning in multimodal models. Advisor: Prof. Qilei Li.
• Model selection feedback — co-authored; submitted to ICLR 2027. Misleading comparison feedback in model selection. Advisor: Prof. Qilei Li.

Human–AI Interaction and LLM Judges

• When Outputs Become Outcomes — first author; submitted to CHI 2027. Distinguishes agreement with an assessment, acceptance of a judge’s authority, and subsequent action in livestream games and academic reviewing; observational, interview, and presentation-comparison methods examine source attribution and human mediation. Advisor: Prof. Qilei Li.

Applied Machine Learning

Healthcare and Clinical Language

• TRACE — accepted at ACM MM 2026. Report supervision during training supports concept-based breast ultrasound diagnosis with image-only inference, in collaboration with Southern Medical University. The work connects BI-RADS concepts, structured clinical descriptions, and concept-edit distillation. Advisors: Prof. Qilei Li and Prof. Zhenyuan Ning.
• Spike and EEG detection: RT-DETR and spatiotemporal graph methods for juvenile and adult rat electrophysiology data, in collaboration with Peking University; algorithm development and signal analysis for symptom studies.
• LLM-based TCM symptom-text augmentation: independent undergraduate honors thesis combining dataset curation, prompting, and fine-tuning. Thesis experiments reduced text length by up to 44.1%, increased terminology tokens 4.4-fold, and achieved diagnostic accuracy of up to 81.3%. Advisor: Prof. Yi Yang.
• Dental orthodontic case retrieval: an ongoing intelligent case-retrieval project in collaboration with Fujian Medical University.

Industrial Inspection and Physical Inverse Design

• FREDNet — ICASSP 2026. Frequency and decomposed-spatial learning with attention fusion for surface-defect detection; contributions include data curation, training, benchmarking, ablations, and cross-domain evaluation on NEU-DET and GC-10, with a reported 4.8% improvement in mAP@0.5. Advisor: Lecturer Dr. Hongjuan Zhang.
• Metasurface inverse design — AIP Advances (2025). CNN–Transformer regression using polarization-conversion spectra and Gaussian-noise augmentation; reported mean MSE of 0.00331. Advisor: Prof. Xiaodong He.

Time-Series Forecasting and Energy Systems

• Artificial Intelligence for Critical Heat Flux in Energy Systems: From Predictive Accuracy to Engineering Use (Shenzhen University RA, 2025.10–2026.5) — review in preparation for submission to Energy. Advisor: Prof. Bo Pang.
• Mamba-In — co-first author; under review at EAAI. Hybrid Mamba and Informer-style ProbSparse attention for long-horizon nuclear sensor forecasting, residual-based fault detection, and signal reconstruction (Shenzhen University RA, 2025.10–2026.5). Advisor: Prof. Bo Pang.
• ReLaMix — arXiv preprint (2026). Residual bottleneck mixing for delayed financial observations, evaluated on second-resolution PAXGUSDT and BTCUSDT using non-overlapping temporal segments and delay-robustness benchmarks. Advisor: Prof. Qilei Li.
• Wind energy prediction — Energies (2025). Modular echo state networks, clustering, and training on turbine data; reported best R² of 0.9905. Advisor: Prof. Zhili Zhao.

Earlier Interdisciplinary Work

• Computational social science and sports media (Sep.–Dec. 2023): AI video generation, Python, Tableau, FreeD 3D, volumetric imaging, and AIGC for studying sports media and player competitiveness. Advisor: Lecturer Dr. Jian Zhan.

Visit history