Ziwei Zhao

Ziway · live my own way

Munich, Germany

ziwei.zhao@tum.de

I am a B.Sc. student in Information Engineering at the Technical University of Munich. In 2025 I was a visiting student at the Gaoling School of Artificial Intelligence, Renmin University of China, where I worked closely with Prof. Yanbiao Ma and Prof. Wenxuan Wang.

My research goal is to make machine-learning models reliable under imperfect data and distribution shift — to measure where they fail, give their predictions statistically valid guarantees, and carry those guarantees into the systems built on top of them. Concretely, I work on fairness and calibration in long-tailed and federated learning, safety evaluation of LLMs and agentic systems, and I am increasingly drawn to causal representation learning and post-training (RLHF/alignment) as the next step. Recurring thread: diagnose the failure, guarantee the fix, and make it survive deployment.

I am applying for Ph.D. programs (Fall 2027) and am open to research collaborations. Please feel free to reach out at ziwei.zhao@tum.de. My full list of papers is on Google Scholar, and PDFs are linked from the publications page.

news

Aug 15, 2026	ScopeBench was accepted to EMNLP 2026 (Main Conference) — a cultural-norm-scope safety benchmark for LLMs. 🎉
Jan 20, 2026	Our paper MessToClean was accepted to ACL 2026 (Main Conference). 🎉
Dec 15, 2025	HeartBridge won Global 3rd Place at the Baidu ERNIE Open-Source Hackathon, where I served as Tech Lead.
Aug 01, 2025	Started a research visit at the Gaoling School of AI, Renmin University of China (RUC GSAI).

selected publications

PR (under review)

Information Density Imbalance in Visual Object Detection

Ziwei Zhao, Yanxi Lu, Yuwei Hu, and 8 more authors

Under review at Pattern Recognition, 2026

Abs arXiv

We define categorized information density as a per-class difficulty signal beyond instance frequency. Its Pearson correlation with AP reaches about -0.72 on Pascal VOC, roughly four times stronger than the instance-count baseline, exposing a long-overlooked bias in long-tailed detector training. A local-to-global covariance merging scheme keeps global information density exact during training and reduces per-class AP variance by up to 45.89% over EFL on LVIS v1.0.
EMNLP 2026

ScopeBench: Evaluating Cultural Norm Scope Awareness in LLMs

Haojun Liang, Ziwei Zhao, Chenyu Li, and 8 more authors

In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP), Main Conference, 2026

Abs PDF

A cultural-scope safety benchmark that treats norm scope as a first-class safety dimension. We build a dual-anchored annotation protocol across 49 countries and 9 cultural clusters (3,323 norms), define three scope error types (Inflation, Deflation, Displacement), and evaluate 21 frontier LLMs across 7 model families.
ACL 2026

MessToClean: Evidence-Grounded Structure-Preserving Reconstruction for Real-World Degraded Exam Paper Images

Jiayi Tuo, Cheng Tang, Zihan Wang, and 8 more authors

In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (ACL), Main Conference, 2026

Abs HTML

An evidence-grounded reconstruction pipeline for degraded exam-paper images. Contributions include benchmark construction and OCR-robustness ablations.