Duanyu Feng 冯端宇

Sichuan University · School of Computer Science
I am a Ph.D. student at Sichuan University, advised by Prof. Wenqiang Lei.
I study representation learning and engineering for language model alignment, graph learning, and AI for science. My current work focuses on RNA language models and structure prediction.
I am currently an exchange student at the Cancer Science Institute of Singapore, National University of Singapore (NUS), working with Prof. Yang Li. Previously, I interned in the Data Research Group at the Beijing Academy of Artificial Intelligence (BAAI), working with Dr. Zheng Zhang on preference data and LLM safety alignment.
Before my Ph.D., I received my B.S. and M.S. degrees in mathematics from Sichuan University.
research experience
- 2025 – PRESENTNUS · Cancer Science Institute of Singapore
Exchange student · RNA language modeling and structure prediction - 2023 – 2024Beijing Academy of Artificial Intelligence
Research intern · Preference data construction, quality filtering, and safety alignment
selected publications
* Equal contribution · † Corresponding author
-
AAAILEGEND: Leveraging Representation Engineering to Annotate Safety Margin for Preference DatasetsAAAI 2025First author
Developed representation-based safety-margin annotation for preference data.
Annotates safety margins using projections along a safety direction in the model’s representation space.
-
CIKMAdvancing Edge Classification through High-Dimensional Causal Modeling of Node-Edge InterplayCIKM 2026 · OralFirst author
Uses balanced edge representations and cross-attention to model node–edge interactions.
-
AAAIAMaPO: Adaptive Margin-attached Preference Optimization for Language Model AlignmentAAAI 2026 · OralResearch guidance and student mentoring
Guided a master's student in developing adaptive-margin preference optimization.
An adaptive margin reallocates learning effort between correctly ranked and misranked preference pairs.
-
NeurIPSHARMONIC: Harnessing LLMs for Tabular Data Synthesis and Privacy ProtectionNeurIPS 2024 · Datasets and BenchmarksResearch guidance and student mentoring
Guided student research on privacy-preserving tabular data synthesis and evaluation.
Uses instruction-tuned LLMs to synthesize tabular data and evaluates utility and privacy leakage.
-
COLINGLAiW: A Chinese Legal Large Language Models BenchmarkCOLING 2025 · Long paperResearch guidance and student mentoring
Guided the development and evaluation of a Chinese legal LLM benchmark.
Evaluates legal language models through tasks organized around legal syllogism.
-
PLoS BiologyDRfold2 is a deep learning-based tool that enables efficient and accurate RNA structure predictionPLoS Biology 2026
Validation, visualization, and manuscript review and editing.
Combines RNA composite language modeling and denoising for structure prediction from a single sequence.