Represent
Alignment · tokenization
- ProtST(ICML 2023 · Oral, top 2%)
- scCello(NeurIPS 2024 · Spotlight, top 2%)
- StructTokenBench(ICML 2025)
I am a Ph.D. student in Computer Science at Mila - Quebec AI Institute and Université de Montréal, advised by Jian Tang. Previously, I earned my B.S. in Computer Science from Peking University, where I worked with Ming Zhang.
My research focuses on multimodal generation, LLM agents and post-training, and biological foundation models. Across these areas, I study out-of-distribution (OOD) generalization to new prompts, tasks, models, and biological settings.
I am currently interning at Tencent Hunyuan with Tianyu Pang. Previously, I interned at Genentech with Aviv Regev and Edward De Brouwer, at Amazon with Zichen Wang and Huzefa Rangwala, and at Intel AI Labs with Hesham Mostafa and Mikhail Galkin.
Before university, I competed in NOIP and NOI algorithm contests, winning NOIP First Prize in both 2016 and 2017 and a bronze medal at NOI 2017.
Alignment · tokenization
Diffusion · flow matching
Tools · VLM feedback
Agent post-training
Learn structure in graphs, proteins, and cells.
Predict cell populations under new treatments.
Use evidence and tools to guide decisions.
Choose tests. Measure outcomes. Update predictions.
Experimental feedback · future direction
* Equal contribution.
Long-video agents combine planning, tool use, visual verification, and repair; execution traces also provide data for LLM post-training.
Uses diffusion Transformers to model cell-response distributions, including responses to unseen drugs or genetic perturbations.
Evidence-grounded LLM reasoning compares related experiments for gene-level perturbation predictions and evaluation on unseen perturbations.
Finds likely protein interaction partners across a whole proteome in minutes by turning residue-level interactions into searchable embeddings.
Compresses long DNA sequences according to genomic regions, helping models learn from longer stretches of DNA with less computation.
Benchmarks discrete 3D tokenization of protein structures and introduces AminoAseed to improve tokenization and downstream performance.
Guides single-cell pretraining with the cell-type ontology to learn gene-expression representations that generalize to new cell types.
Pretrains a single graph foundation model on relational patterns so it can reason over unseen knowledge graphs.
Cross-modal pretraining aligns protein sequences with biomedical text for zero-shot classification and text-guided protein retrieval.
Learns which paths to search first, enabling fast and interpretable reasoning across million-scale knowledge graphs.
Provides an open-source PyTorch library of datasets, models, and benchmarks for molecules, proteins, and biomedical knowledge graphs.