Hi, I’m Jiaqi Sun, currently a PhD student from Carnegie Mellon University working with the CMU CLeaR Group and advised by Prof. Kun Zhang and Prof. Clark Glymour. Previously, I completed my Master’s at Tsinghua University in 2023, when I was advised by Prof. Yujiu Yang, and obtained my bachelor’s degree from Sun Yat-sen University in 2020.
My ultimate goal is to build learning models that have direct interpretability and can continually evolve from well-constructed representation systems. To borrow insights from current optimal solutions in creatures like humans, in parallel, I am trying to understand brain activity from various neuronal signals. For both branches, I am grounded in causal representation learning theory and assume there is a data-generative process responsible for producing responses or observational signals. Beyond that, I also care about how AI impacts the human brain and society before it becomes a fully transparent tool.
Contact me at sajqavril at gmail dot com, google scholar
** denotes equal contribution.*
Towards interpretability, continually evolving learning models:
Jiaqi Sun, Boyang Sun, Rasmy M. H., Xiangchen Song, Kun Zhang. MoRe: Modular Representations for Principled Continual Representation Learning on Sequential Data. Preprint (2026).
We proposed learning hierarchical, modular representations for sequential data that distinguish fundamental knowledge from more domain-specific knowledge. This modularity naturally supports a principled adaptation method, so the knowledge/representation can evolve with selective updates.
Kun Zhang*, Jiaqi Sun*, Yiqing Li, Ignavier Ng, Namrata Deka, Shaoan Xie. SEDGE: Structural Extrapolated Data Generation. Preprint (2026).
We derived a theory of reliable extrapolated generation: under what data distribution conditions and data structures reliable extrapolative generation can be conducted by the proposed algorithm. This draws the practical boundary for the maximum utilization of training data.
Xiangchen Song* Jiaqi Sun*, Zijian Li, Yujia Zheng, Kun Zhang. LLM Interpretability with Identifiable Temporal-Instantane`ous Representation. NeurIPS (2025).
We extended conventional Sparse Autoencoders (SAEs) to model temporal dependencies between features with identifiability guarantees. It works efficiently on real-world LLM activations and reveals interpretable features and their relations.
Jiaqi Sun, Yujia Zheng, Xinshuai Dong, Haoyue Dai, Kun Zhang. Type Information-Assisted Self-Supervised Knowledge Graph Denoising. AISTATS (2025).
We proposed obtaining the most compact set of relations from any knowledge graph, guided by type information, to assist compression, denoising, and completion. Working on a journal extension.
During my Master’s, I was focusing on explaining the behavior of Graph Neural Networks (GNNs) before I realized that I am more generally caring about the causal mechanisms of observations.
Jiaqi Sun, Lin Zhang, Guangyi Chen, Kun Zhang, Peng Xu, Yujiu Yang. Feature Extension for Graph Neural Networks. ICML (2023).
We suggested a unified feature space perspective to understand the varying performance of different GNNs on different graphs and, accordingly, proposed an efficient and effective linear GNN.
Jiaqi Sun, Lin Zhang, Shenglin Zhao, Yujiu Yang. Improving Your Graph Neural Networks: A High-Frequency Booster. ICDMW (2022).
We proposed a flexible high-pass filter plugin for GNNs to enhance performance on heterophilic graphs using graph signal processing.
I have unexpectedly experienced severe pmdd for months, from which, for the first time, I learned how to listen to my body and respect that everyone, including myself, is just a human being. We might be more sensitive at times, especially during cycle changes or particular occasions—all of them are absolutely fine. Before pushing yourself to the limit, it is okay to slow down, reflect, and build a sustainable schedule and lifestyle. I am lucky enough to be supported, understood, and helped by many people around me. You can shine at any time if you are on your own pace.
If you’ve experienced significant stress and an abnormal mental or physical reaction that has never happened before, it is worth pausing and thinking about why, and there is a lot of help available around you, including me. You are never alone; we just have our special ways to go. Everyone does.
I paint with oil pastels during spare time.
