Hong Kong Baptist University under the supervision of One paper TSTUDIO has been accepted by ICDE 2027!
Discovers semantic TSketch tokens that shorten long time-series inputs, accelerating Transformer training by 2–120× while maintaining or improving classification accuracy.
Guozhong Li
Postdoctoral Research Fellow · KAUST
LLM Systems · Temporal & Scientific Data · Vector Databases
I develop efficient and trustworthy AI systems for large language models and high-dimensional data, spanning LLM inference, temporal representation learning, scientific-data compression, and privacy-preserving vector retrieval.
At KAUST, I am fortunate to be advised by Prof. Panos Kalnis.
Prior to that, I completed my Ph.D. in Computer Science at Hong Kong Baptist University under the supervision of Prof. Byron Choi and Prof. Jianliang Xu. I was also a member of the HKBU Database Group.
My dissertation was Shapelet Discovery for Time Series Analysis.
🔥 News
Our proposal has been approved for six months of research compute support through the Google TPU Research Cloud (TRC) program.
Google TPU Research Cloud (TRC)Building on LLMComp, the project will explore TPU-trained foundation models for error-bounded scientific data compression.
Our preprint MentorPulse is now available on arXiv.
Refreshes cross-model latent guidance during decoding to sustain small-model performance in long-form generation.
Our preprint KVDiagnosis is now available on arXiv.
Provides a diagnostic benchmark for identifying when and why KV-cache compression fails in long-context LLMs.
One paper HTSA has been accepted by KDD 2026!
Selects representative subgraphs to summarize hierarchical time series while preserving structural and temporal patterns.
One paper GraphComp has been accepted by TKDE!
Uses temporal graph autoencoders to compress scientific data under user-defined point-wise error bounds.
Our paper LLMComp was published at IEEE BigData 2025.
Recasts scientific data compression as LLM-based top-k prediction while preserving strict error bounds.
🔍 Research
Across these areas, I combine learning, indexing, and systems techniques to balance accuracy, scalability, and trustworthiness.
Temporal Representation Learning
Shape-based, graph-driven, and transformer representations for temporal classification, forecasting, and anomaly detection.
Selected work:ShapeNet: AAAI'21, AutoShape: arXiv, IPS: ICDE'22, BSPCover: TKDE'22,SVP-T: AAAI'23, DARKER: PVLDB'24, Hierarchical TS Abstraction: KDD'26, TSTUDIO: ICDE'27 Ongoing: Learned graph structures for time-series classification; symbolic and token-based transformers; multimodal and foundation models for time series; hierarchical time-series anomaly detection.Error-Bounded Compression
Graph, neural, and language-modeling frameworks that preserve user-defined error guarantees.
Selected work: GraphComp: TKDE'26, LLMComp: BigData'25 Ongoing: DNN4LSDC.Efficient LLM Inference
Semantic context selection, KV-cache optimization, runtime attention control, and coordinated multi-model reasoning.
Selected work: CHESS: arXiv, ART: arXiv, KVDiagnosis: arXiv, MentorPulse: arXiv Ongoing: KV-cache orchestration and adaptive multi-model coordination.Privacy-Preserving Vector Databases
Secure indexing and approximate retrieval over sensitive, high-dimensional data.
Selected work: leSAX: ICDE'25 Ongoing: Privacy-preserving approximate nearest-neighbor search over high-dimensional data.☁️ Research Support
Google TPU Research Cloud (TRC) Six-month research compute grant.
🎖 Honors & Awards
- Research Performance Award · HKBU
- Travelling Allowance for Postgraduate Overseas Attachment Programme · HKBU
- Excellent Teaching Assistant Performance Award · HKBU
- Teaching Assistant Performance Award · HKBU
Earlier honors
- Outstanding Graduate Student · UESTC
- IBM-CSC Scholarship
- National Scholarship