About Me
I am a first-year MPhil student in Artificial Intelligence at the Information Hub, HKUST (Guangzhou), advised by Prof. Hui Xiong and Prof. Ying Sun. I received my B.S. in Physics with a minor in Computer Science from Shandong University in 2026 — a background that keeps me interested in how models work inside, not only what they produce.
My research focuses on post-training for large language models: shaping the computation behind answers, deciding which examples are worth learning from, and turning interpretability signals into practical training tools. I am always happy to chat about research — feel free to reach out by email.
Research Interests
- Process-level self-distillationsupervising how models reason in hidden space, not only what they output
- Training-data selectionchoosing high-value examples with interpretable coverage signals and verifiers
- SAE interpretability for trainingusing sparse features for steering, continual learning, and evaluation
- Multimodal alignment & AI for Sciencefrom molecular editing to reasoning over raw physical measurements
News
Started as an MPhil student in Artificial Intelligence at HKUST (Guangzhou).
Released a survey on large language models for molecular science on ChemRxiv (equal contribution).
Released EMRB on arXiv (under review at KDD 2027 DB Track).
Submitted PHF, Self-Correcting Bradley–Terry, and RidgeSteer to AAAI 2027.
Released PHF on arXiv.
Submitted IRDS, SLIM, and SAE-FD to EMNLP 2026 (preprints available).
Publications
-
Large Language Models for Molecular Science: A Survey on Representation and Cognition
A survey of how large language models represent and reason about molecular science, covering representation learning and cognitive perspectives. Equal contribution.
-
EMRB: A Multi-Level Benchmark for Evaluating LLM Reasoning over Raw Electromagnetic Signals
A multi-level benchmark asking whether LLMs can reason over raw electromagnetic I/Q signals by writing and running code, from low-level measurement to system design.
-
PHF: Privileged Hidden Flow for On-Policy Self-Distillation
Distills how a privileged EMA teacher’s hidden states move along the student’s own rollouts — aligning transition directions and trajectory geometry on top of output-level on-policy self-distillation.
-
Self-Correcting Bradley–Terry: Turning Ranking Residuals into a Bias Detector for LLM-as-a-Judge
Turns the residuals of a Bradley–Terry fit over pairwise rankings into a self-contained detector of systematic bias in LLM-as-a-judge pipelines.
-
RidgeSteer: Ridge-Regularized Activation Steering for Large Language Models
Stabilizes activation steering with ridge regularization, trading raw steering strength for reliability across prompts and layers.
-
IRDS: Interpretable RLVR Data Selection via Verifier-Coupled Sparse Autoencoder Coverage
Selects RLVR training data by coupling sparse-autoencoder feature coverage of each example with verifier signals, making the selection criterion itself interpretable.
-
SLIM: Sparse Latent Steering for Interpretable and Property-Directed LLM-Based Molecular Editing
Steers molecule-editing LLMs in a sparse latent space, so each edit direction is interpretable and tied to a target property.
-
SAE-FD: Sparse Autoencoder Feature Distillation for Continual Learning of Large Language Models
Distills sparse-autoencoder features from earlier tasks to reduce catastrophic forgetting in continual learning of LLMs.
Experience
-
Oct 2025 – present
HKUST (Guangzhou) — Research Assistant
- Process-level self-distillation for LLM mathematical reasoning, using hidden-state transition signals to shape reasoning representations.
- Interpretable data selection for RLVR: sparse-autoencoder coverage combined with verifier signals to pick high-value training examples.
-
Jun – Oct 2025
SUSTech — Visiting Student
- CLIP/SigLIP post-training with curated instruction and caption data for image–text alignment.
- Evaluation on retrieval and zero-shot multimodal benchmarks; analysis of data quality vs. performance.
-
Apr – Jun 2025
CUHK — Research Assistant
- Generative modeling of disordered crystals with space-group and Wyckoff tokenization constraints.
- Data-processing, structure-validity, and latent-space analysis pipelines linking structure to properties.
Education
-
Sep 2026 –
HKUST (Guangzhou)
MPhil in Artificial Intelligence, Information Hub
-
2022 – 2026
Shandong University
B.S. in Physics, minor in Computer Science and Technology
Awards & Honors
First Prize (Provincial), China International College Students’ Innovation Competition (“Internet+”)
Third Prize (National), China Undergraduate Physics Experiment Competition
First Prize (Provincial), Shandong Undergraduate Physics Experiment Competition
Second Prize (Provincial), China Undergraduate Mathematical Contest in Modeling
Third Prize (Provincial), National Competition on Energy Conservation & Emission Reduction
Third Prize (Provincial), “Challenge Cup” Extracurricular Academic Science & Technology Competition
Third Prize (Provincial), “Datang Cup” ICT Competition
Principal Investigator, Shandong University Student Innovation Fund project
First-, Second-, Third-Class, and Special-Talent Scholarships, Shandong University