Zhuohao Yu

About Me

Hello, I am Zhuohao Yu (于 倬浩), currently a second-year Master’s student at Peking University, advised by Prof. Wei Ye and Prof. Shikun Zhang. My research interest lies in natural language processing.

Beyond academia, I have a strong background in competitive programming, where I won Gold medal in the ACM-ICPC regionals and Silver medal in the National Olympiad in Informatics (NOI).

Education

• 2023.09 - Present, MSc, Software Engineering, Peking University.
• 2019.09 - 2023.06, BSc, Computer Science, Renmin University of China.

Research Interests

My research focuses on advancing interpretable, reliable, self-evolving LLMs through autonomous evaluation, alignment and reasoning, with a commitment to creating trustworthy, practical, open-source AI systems.

This work follows a progressive pipeline: developing robust evaluation methodologies, leveraging these insights for self-improvement, all built upon efficient and trustworthy infrastructures.

• Autonomous evaluation of language models. As LLMs surpass human expertise in specialized domains, evaluation becomes profoundly challenging. When models possess knowledge beyond human validators, who becomes the arbiter of truth? How can evaluation frameworks evolve alongside increasingly capable models? What distinguishes a model that truly understands from one that merely memorizes?

  Related Works: KIEval (ACL 2024), PandaLM (ICLR 2024), FreeEval (EMNLP 2024).

• Self-improving and reasoning LLMs. The next frontier lies in creating models that can leverage autonomous rewards to enhance their own capabilities, both during training and inference. How can we convert assessment into actionable learning that preserves model integrity? What role does reasoning play in enabling models to refine their own capabilities? What constraints prevent optimization for superficial metrics rather than meaningful capabilities?

  Related Works: ORPS (Preprint), Supervised Knowledge… (ICLR 2024).

• Trustworthy open-source AI systems. Realizing AI’s potential demands infrastructure that is powerful, accessible, and responsible. How can we democratize access while establishing safeguards against misuse? What architectures allow transparency without sacrificing efficiency? How might attribution mechanisms maintain accountability within collaborative ecosystems?

  Related Works: SAEMark (Preprint), FreeEval (EMNLP 2024), PaperLens (Try it).

Publications

Fun Projects

Services

Conference Reviewers

• ICLR
• ACL / EMNLP / NAACL (Regular ARR Reviewer)
• AISTATS

Journal Reviewers

• ACM Transactions on Intelligent Systems and Technology

京ICP备2024044818号