Yiming Qin

I am a computer science PhD student at the École Polytechnique Fédérale de Lausanne (EPFL). I am currently working on graph generation and have an interest in drug discovery and conditional generation. I am fortunate to be advised by Prof. Pascal Frossard.

Prior to my PhD, I completed my Master's degree at the at Shanghai Jiao Tong University (SJTU), where I worked with Prof. Ya Zhang, Prof. Yanfeng Wang, Prof. Jiaochao Yao, and Dr. Huangjie Zheng. At the same time, I earned the double degree (title of Ingénieur) from École polytechnique (X) on applied mathematics, finished with an internship at École Normale Supérieure (ENS), Paris under the supervision of Dr. Lisa Bedin and Prof. Auguste Genovesio. Before that, I completed my Bachelor's degree a B.E. in Information Engineering and a B.A. in French at SJTU, within the SPEIT.

Email  /  GitHub  /  Google Scholar  /  LinkedIn

News

Research

My research focuses on generative models and graphs. Specifically, I am interested in controlling generative models, such as model editing and conditional generation, the intersection of autoregressive and diffusion language models, and applying graph generative models to diverse domains, such as drug discovery.

DBMol: Design of High-Affinity, Target-Specific Small Molecules through Structure Prediction Models

DBMol is a docking-dataset-free framework that uses structure prediction models to guide de novo small-molecule design, achieving strong affinity and pocket specificity without reinforcement learning or docking supervision. This project applies 2D graph generation for small-molecule design.

Fixed-Point Masked Generative Modeling

We introduce CoFRe, a fixed-point framework for masked generative models with a looped architecture that reuses shared parameters, reducing parameters, training cost, and memory usage while improving low-budget generation across text and image modalities.

Efficient Edge-aware Attention Network for Graph Generation

We propose a kernel-friendly edge-aware attention layer for graph generation that is directly compatible with PyTorch FlexAttention, reducing GPU memory by around 65%. This work offers a complementary angle to SparseDiff for scalable graph generation.

Generating Directed Graphs with Dual Attention and Asymmetric Encoding

We propose DIRECTO, a discrete flow matching model with direction-aware encodings and dual-attention, plus new benchmarks, for better directed graph generation. It extends ideas from DeFoG to directed graphs and shows their robustness in a new setting.

DeFoG: Discrete Flow Matching for Graph Generation

Graph generation models often face limitations in sampling efficiency and flexibility due to tightly coupled training and sampling stages. We introduce DeFoG, a discrete flow matching framework that disentangles these stages, improving both efficiency and performance over existing diffusion models.

SparseDiff: Sparse Discrete Diffusion for Scalable Graph Generation

Generative models for graphs struggle with scalability due to predicting interactions for all node pairs. We introduce SparseDiff, a denoising diffusion model that leverages sparsity to overcome this. It reformulates diffusion, denoising, and the model architecture around sparse graph representations.

Class-balancing Diffusion Models

We observe significant performance degradation on tail classes in both diversity and fidelity, and propose Class-Balancing Diffusion Models (CBDM) that are trained with a distribution adjustment regularizer as a solution.

Supervised projects

Here are some research projects that I have supervised. Open projects (for EPFL students) can be found here. Feel free to get in touch.

Balancing Symmetry and Efficiency in Graph Flow Matching

We study the trade-off between symmetry and efficiency in graph flow matching, showing that properly modulating equivariance can accelerate convergence while delaying overfitting.

On the Role of Structure in Hierarchical Graph Neural Networks

Hierarchical GNNs still don’t outperform flat ones overall but recover better from structural perturbations.

Projects

Here are some projects that I have worked on.

Learning Cell Dynamic Features for Unlabeled Video-Microscopies from Synthetic Data

We design a multi-task model with two heads: one trained on synthetic mitosis videos (SynDiv) using supervised/self-supervised tasks, and one trained on real microscopy videos (LSP) using self-supervised tasks. Joint training improves the learned representations, yielding better correlations with cell counts, division events, and genetic perturbations than models trained on SynDiv or LSP alone.

Design and source code from Leonid Keselman and Jon Barron's website