Projects

Research and project experience, grouped by theme. Papers are listed on the Publications page.

My research spans model evaluation, reliable LLM reasoning, and human–AI collaboration, and connects them with applied AI and machine learning.

Logos on each entry: the rightmost is my home institution; any to its left are the main collaborating institutions.

AI Evaluation and Analysis

  • ExposureBench: Knowing When to Update — Evaluates when web evidence selection should change a model’s probability and when disclosure makes the same policy change redundant. Crosses selection policies with information disclosure using exact finite-population targets and TREC DL retrieval with BM25 and Qwen reranking. Measures probability accuracy, informative updates, and conditional policy drift separately, using repeated calls and decision replay to assess noise and downstream consequences. Advisor: Prof. Qilei Li
  • Choice2Seq — arXiv 2026 Tests whether a model that selects the correct complete answer can also construct it through local sequential decisions. Sequence Composition Rate tracks that retention under serial construction: Jev keeps 33.0% and 2.1% of Native-correct arithmetic and lexical items despite 97.1% and 99.5% selection accuracy. Advisor: Prof. Qilei Li
  • Which Differences Matter? — arXiv 2026 Controlled tests of which rule and information changes guide decision-model answers: transfer on semantically separated rules, joint reliability on crossed rule–state tests, and directed use of table values versus mappings on Jev. Advisor: Prof. Qilei Li
  • When Does Agreement Add Information? — Sensor and document verification with exact reference probabilities, varying copying versus independence so agreeing answers should update differently, then scoring both final probability accuracy and subsequent update correctness. Advisor: Prof. Qilei Li
  • Reasoning after Unlearning — Evaluates original, unlearned, and target-excluded retrained Llama checkpoints on TOFU, separating context effects, answer exposure, and use of supplied answer information with matched controls, masking, and identity substitution. Advisor: Prof. Qilei Li
  • Multimodal unlearning — Auxiliary supervision for unlearning in multimodal models, using extra training signals while target information is being removed. Advisor: Prof. Qilei Li
  • Model selection feedback — Misleading comparison feedback in model selection, studying how distorted comparisons affect which model is chosen. Advisor: Prof. Qilei Li
  • When Outputs Become Outcomes — Distinguishes agreement with an assessment, acceptance of a judge’s authority, and subsequent action in livestream games and academic reviewing, using observational, interview, and presentation-comparison methods to examine source attribution and human mediation. Advisor: Prof. Qilei Li

Applied Machine Learning

Clinical and Dental

  • Oral Benchmark — A benchmark built from dental and orthodontic cases.
  • Dental orthodontic case retrieval — An ongoing case-retrieval project for orthodontic records, with Fujian Medical University.
  • TRACE — ACM MM 2026 Uses report supervision during training so that concept-based breast-ultrasound diagnosis needs only the image at inference, linking BI-RADS concepts, structured clinical descriptions, and concept-edit distillation; joint work with Southern Medical University. Advisors: Prof. Qilei Li & Prof. Zhenyuan Ning
  • Spike and EEG detection — RT-DETR and spatiotemporal graph methods for detecting spikes and EEG events in juvenile and adult rat electrophysiology, developed with Peking University for algorithm work and signal analysis in symptom studies.
  • LLM-based symptom-text augmentation for Traditional Chinese Medicine (TCM) — Independent undergraduate honors thesis covering dataset curation, prompting, and fine-tuning; experiments cut text length by up to 44.1%, increased terminology tokens 4.4-fold, and reached up to 81.3% diagnostic accuracy. Advisor: Prof. Yi Yang

Inspection and Inverse Design

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

Forecasting and Energy

  • Artificial Intelligence for Critical Heat Flux in Energy Systems: From Predictive Accuracy to Engineering Use (Shenzhen University, Oct. 2025 – present) — Review proposal sent; now contacting Energy for an invitation to submit the survey. Advisor: Prof. Bo Pang
  • Mamba-In (Research assistant, College of Physics and Optoelectronic Engineering, Shenzhen University, Oct. 2025 – May 2026) — Combines Mamba with Informer-style ProbSparse attention for long-horizon forecasting, residual-based fault detection, and signal reconstruction on high-reliability nuclear-sensor monitoring streams. Advisor: Prof. Bo Pang
  • ReLaMix — arXiv 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 with clustering, trained on turbine data; best reported R² of 0.9905. Advisor: Prof. Zhili Zhao

Federated Learning

  • FedSocratic — Verifiable questioning for federated language models, so questioning steps in the federated setting can be checked. Advisors: Prof. Qilei Li & Prof. David Camacho
  • After Cooperation Is Learned — arXiv 2026 Gradient routing and optimizer-dependent maintenance of learned cooperation once multi-agent reinforcement learning has already acquired it. Advisor: Prof. Qilei Li
  • MOSAIC — arXiv 2026 Margin-oriented semantic–appearance interaction correction for personalized federated vision–language models, going beyond domain-level adaptation. Advisor: Prof. Qilei Li
  • FedSAP — arXiv 2026 Structured adaptive partitioning for federated learning on multi-domain heterogeneous edge devices. Advisor: Prof. Qilei Li
  • Data heterogeneity survey — Expert Systems 2026 Led the survey framework on federated learning under non-IID and heterogeneous data, covering healthcare, finance, and IoT, and designed frequency-adaptive semantic distillation with semantic topology and dynamic frequency-based weighting. Advisors: Prof. Qilei Li & Prof. David Camacho

Early Research in Social Computing

  • Social computing in sports media (Sep. – Dec. 2023) — Studied sports media and player competitiveness using AI video generation, AIGC, FreeD 3D, volumetric imaging, Python, and Tableau. Advisor: Lecturer Dr. Jian Zhan