🔬 Research

Hanzhi’s research primarily focuses on ⚡ Efficient AI Systems and 🤖 Reliable and Responsible AI.

⚡ Efficient AI Systems

Her primary research direction spans hardware-aware LLM inference, sparse and mixed-precision computation, post-training quantization, GPU optimization, and HPC. She conducts large-scale GPU experiments on ALCF Polaris and TACC Lonestar6.

DAM: Adaptive Sparse Attention

DAM: Dynamic Attention Mask for Long-Context Large Language Model Inference Acceleration develops adaptive masks at the level of individual attention maps, preserving different attention patterns across layers and heads. The work explores efficient long-context inference without fine-tuning or predefined mask structures.

ACL Findings 2025 · Paper · Code

🤖 Reliable and Responsible AI

Poly-FEVER: Multilingual Fact Verification

Poly-FEVER provides 77,973 labeled factual claims across 11 languages for studying fact verification and hallucination detection in large language models. Hanzhi led the development of this benchmark to support research on reliable AI across languages.

EMNLP Findings 2026 · Paper · Dataset

VQS-4K: Visual Query Segmentation in the Wild

Towards Visual Query Segmentation in the Wild introduces a benchmark with 4,111 videos and more than 1.3 million frames. Hanzhi co-developed VQS-4K to study visual query segmentation in real-world video.

NeurIPS 2026 · Co-first author, with Bing Fan and Minghao Li (equal contribution) · Paper

D2SO: Safety-Critical Perception

D2SO investigates the detection of distant and small objects in vision-based autonomous vehicle systems, supporting safer perception in challenging driving environments.

IEEE MOST 2025 · Paper

Hallucination Analysis and Scientific Data

Her collaborative work includes HALO, which studies hallucination analysis and retrieval-augmented context for clinical decision making; Harmful Factuality, which examines how factual corrections can produce outputs that are true but unfaithful to the source; and ByteScience, which bridges scientific literature and structured data with language models.

📚 Explore all peer-reviewed publications

🤝 Grants, Computing Allocations & Collaborations

Amazon Research Awards: TileRoute-Trainium · 2026–present

Hanzhi serves as lead student researcher and proposal co-writer on TileRoute-Trainium, an Amazon Research Awards project led by Dr. Yunhe Feng at UNT. In collaboration with Qian Qu, an Amazon scientist, she investigates efficient, mixed-precision LLM inference using AWS Trainium accelerators.

Oak Ridge National Laboratory: AAIMS Group · 2026–present

As a Research Collaborator in the Analytics and AI Methods at Scale (AAIMS) Group, she works with Dr. Emily Herron and Dr. Aditya Kashi in Dr. Feiyi Wang’s group on AI-based 3D boundary-to-domain operator learning for partial differential equations. The ongoing collaboration is conducted remotely.

ALCF Polaris Computing Allocations · 2024–present

Hanzhi serves as PI for computing allocations on ALCF Polaris, with 5,000 node-hours per allocation, for the project Large Language Models’ Hallucination Detection and Mitigation Ability across Multi Languages.

Argonne Training Program on Extreme-Scale Computing · Summer 2024

She was competitively selected for ATPESC 2024, an intensive two-week program covering GPU computing, parallel programming models, and performance optimization on leadership-class supercomputers.

Her research training resources include a TACC Lonestar6 HPC tutorial and a Docker setup tutorial, developed to help junior researchers conduct experiments independently.