Not X, but When: Diagnosing a Mandarin LLM Writing Tic through Contextual Appropriateness
A pragmatics-grounded framework for judging when the Mandarin “not X, but Y” construction is contextually appropriate, and why LLMs overuse it.
Research Archive
Publications
Papers and preprints grouped by year, with direct links to papers, code, datasets, and project materials.
2026
Guiding LLM Post-training Data Engineering with Model Internals from Sparse Autoencoders
A sparse autoencoder-guided framework for using model internals to engineer post-training data for LLM reinforcement learning.
HistLens: Mapping Idea Change across Concepts and Corpora
Accepted at ACL 2026 Main Conference. Code available on GitHub.
2025
DeepEddy: High-quality Fast Eddy Current and Bulk Motion Correction Using Deep Learning-based Image Synthesis and Co-registration
Deep learning-based image synthesis and co-registration for fast eddy current and bulk motion correction in diffusion MRI. Magna Cum Laude Merit Award.
UDTI: High-fidelity Ultra-fast Diffusion Tensor Imaging in Stroke Patients Using Transfer Learning
A transfer-learning approach that shortens diffusion tensor imaging in stroke patients from 9.7 minutes to 1 minute while preserving image quality and tractography.
LinguaLens: Towards Interpreting Linguistic Mechanisms of Large Language Models via Sparse Auto-Encoder
Accepted at EMNLP 2025 Main Conference. An earlier version appeared at the XLLM Workshop @ ACL 2025.