Zesheng Ye
Greetings! I am a National Intelligence Postdoctoral Fellow at the TMLR Group, Faculty of Engineering and Information Technology, University of Melbourne, hosted by Dr. Feng Liu. I did my Ph.D. studies at the University of New South Wales, feeling fortunate to have been supervised by Prof. Lina Yao.
I study how to make machine learning systems trustworthy across their lifecycle, with a particular emphasis on:
- Reliable prediction (RP): knowing what models don't know under data scarcity
- Reliable adaptation (RA): reusing foundation models for new tasks effectively
- Reliable evaluation (RE): certifying performance and safety claims of deployed models
My long-term goal is to build ML systems whose reliability can be quantified, justified, and verified, such that AI can be safely deployed for the general public with guaranteed confidence.
News
- Apr 2026Two preprints on defending against model/data laundering (RE), for distilled diffusion models (ICML R2-FM Workshop 2025) under model laundering and LLMs (ICLR Data-FM Workshop 2026) under data laundering.
- Jan 2026One paper on RA accepted to ICLR 2026, regarding class-aware zero-shot prompt reweighting for VLMs.
- Jan 2026Co-organize a tutorial on Neural Network Reprogrammability at AAAI 2026, Singapore, with IBM Research AI.
- Nov 2025Grateful to receive the DAAD AINet Postdoctoral Fellowship.
- Sep 2025Grateful to receive the RecSys 2025 Distinguished PC Member award.
- Jul 2025Grateful to receive the National Intelligence Postdoctoral Grant.
- Jun 2025Grateful to receive the ICML 2025 Top Reviewer award (top 2%).
- May 2025Two papers accepted to ICML 2025, regarding decoupling the roles of visual prompts for VLMs (on RA) and input-dependent calibration for adversarial purification (on RE).