Exposing Privacy Risks in GraphRAG
First empirical study of GraphRAG privacy risk — up to 73.6% entity leakage per query.
Read MoreMSCS @ UC San Diego
UCSD MSCS · seeking Summer 2027 ML Engineer internship · GNNs, LLMs, retrieval
I'm Jiale Liu, a current Master of Computer Science student at UC San Diego (Sep 2026–Jun 2028). I earned my B.S. in Computer Science from Huazhong Agricultural University. I work at the intersection of graph neural networks (GNNs), large language models (LLMs), and retrieval, and I'm seeking a Summer 2027 ML Engineer internship.
I interned at vivo AI Lab on BlueLM pre-training data, worked with Prof. Suhang Wang at Penn State on GraphRAG privacy, and conducted signed-graph interpretability research with Prof. Zeyu Zhang.
Beyond academics, I enjoy running, table tennis, and traveling. I've visited Japan, Thailand, Vietnam, Xinjiang, Tibet, and many other unforgettable destinations.
University of California, San Diego
Huazhong Agricultural University
PyTorch, Hugging Face, vLLM, Python, Docker
FAISS, RAG, GraphRAG, retrieval / recommendation
GNNs, signed graphs, Graph Transformers, knowledge graphs
C++ / SQL, large-scale data pipelines
Industry internship and research appointments.
vivo AI Lab · Shenzhen
BlueLM pre-training data: near-deduplication with MinHash LSH Ensemble, an ASR pipeline on vLLM / Qwen, and large-scale data work with tracked quality and throughput metrics.
Penn State · Prof. Suhang Wang's Lab
First empirical study of privacy leakage in GraphRAG (ACL 2026 Findings): up to 73.6% entity leakage and 74.0% relationship leakage per query.
Huazhong Agricultural University · Prof. Zeyu Zhang
Signed-graph interpretability: SE-SGformer (AAAI 2025, +73.1% explainability) and SSEFormer (Neurocomputing, +17.0% AUC on Amazon-Music).
Exploring interpretability, privacy, and trustworthiness in AI systems.
First empirical study of GraphRAG privacy risk — up to 73.6% entity leakage per query.
Read More
LLMs translate graph structure into text for end-to-end interpretability.
Read More
SE-SGformer: up to +73.1% explainability over signed-graph baselines.
Read More
+17.0% AUC over SIGformer on Amazon-Music.
Read MoreRetrieval systems and applied machine learning.
Hybrid feature extraction with FAISS-powered millisecond-level similarity search.
Read More
Automated extraction from academic documents with interactive graph visualization.
Read MoreSome things people often ask me.
Interested in collaboration or just want to chat? Feel free to reach out.