Hi there! I am an undergraduate researcher at Jinan University, working at the intersection of evaluation-centred trustworthy AI and data-centric LLM systems. My research asks how we can make modern models more measurable, interpretable, and efficient.

My work spans information-theoretic evaluation, causal model auditing, mechanistic interpretability, efficient instruction-data selection, and generative models for computer vision. I am currently seeking fully funded PhD opportunities and welcome contact from prospective supervisors and collaborators.

Research Interests

AI Evaluation and Model Auditing

Designing model-agnostic protocols that test whether benchmarks and models use the information they are expected to use.

Mechanistic Interpretability

Studying how learned representations encode task-relevant features, with sparse autoencoders and causal interventions.

Data-centric AI and Efficient LLM Systems

Building mathematically grounded pipelines for instruction-data selection, efficient training, and reliable model development.

Trustworthy and Causal ML

Using causal, information-theoretic, and probabilistic analysis to diagnose structure, uncertainty, and failure modes.

News

• Oct 5, 2026: Submitted CDIS: Training-Free Data Selection for Instruction Tuning with Compression Ratio and Token Entropy to ACL ARR 2026 October.
• Sep 29, 2026: Submitted Residualizing DKT with Symbolic Sequence Priors for Knowledge Tracing to AISTATS 2027.
• Sep 25, 2026: An Information-Theoretic Evaluation Framework for Benchmark and Model Diagnosis in Knowledge Tracing was accepted as a poster at NeurIPS 2026.
• Sep 17, 2026: Submitted When Target Accuracy Hides Semantic Interface Drift in Vision–Language Adaptation to ICLR 2027.
• Apr 28, 2026: Submitted An Information-Theoretic Evaluation Framework for Benchmark and Model Diagnosis in Knowledge Tracing to NeurIPS 2026.
• May 2025: Joined Prof. Mingliang Hou’s group at Jinan University as an RA.
• Nov 2024: Joined Prof. Hao Chen’s group at Macau University of Science and Technology as an RA.

Selected Publications

^* Equal contribution / co-first author.

Avatar image: official character artwork from the TV anime character page.