Yafu Li (李雅夫)

I am a researcher at Shanghai AI Lab, working under the supervision of Prof. Yu Cheng. I earned my PhD through a joint program between Zhejiang University and Westlake University, advised by Prof. Yue Zhang.

I received my Bachelor's degree from Wuhan University, followed by a Master's degree from the University of Edinburgh, where I was supervised by Prof. Alex Lascarides. During my PhD, I interned at Tencent AI Lab and collaborated closely with Dr. Leyang Cui and Dr. Wei Bi. Prior to that, I worked as an NLP researcher at Huawei Noah's Ark Lab, under the mentorship of Dr. Liangyou Li and Prof. Qun Liu.

Email  /  Google Scholar  /  Twitter  /  Github

We are looking for interns and joint PhD candidates (with THU, PKU, SJTU, FDU, etc.) to work on cutting-edge research in large language models. Our focus areas include zero reinforcement learning (e.g., R1-zero), test-time scaling, and trustworthy AI. If you are interested, please feel free to contact me at yafuly@gmail.com.

My research focuses on reasoning and inference scaling, trustworthy AI and machine translation. *: equal contributions. †: project lead or corresponding author.

A benchmark to systematically evaluate and analyze the instruction-following ability of large reasoning models.

A RL framework to boost reasoning performance using off-policy guidance.

We survey recent efforts to improve reasoning efficiency in Large Reasoning Models, highlighting inefficiency patterns in long reasoning traces and discussing solutions across the model lifecycle from pretraining to inference.

We present Test-Time Preference Optimization (TPO) that aligns LLMs during inference, surpassing strong baselines aligned during training.

We present Aggregation Fine-Tuning (AFT) where the model learns to aggregate multiple drafts into a single answer.

We propose Mixture-of-Search-Agents (MoSA), a test-time framework that coordinates multiple LLMs via Monte Carlo Tree Search to improve complex reasoning by aggregating diverse search trajectories.

LLMs exhibit stable paraphrasing cycles revealing limits in linguistic diversity through dynamical systems analysis.

We present a comprehensive benchmark dataset designed to assess the proficiency of machine-generated text detectors amidst real-world scenarios.

A survey of hallucination in LLMs.

A systematic study of the origin of translationese in LLMs and mitigation methods.

We propose a syntax-guided generation schema, which generates the sequence guided by a constituency parse tree in a top-down direction.

We propose multi-granularity optimization for non-autoregressive translation, which collects model behaviors on translation segments of various granularities and integrates feedback for backpropagation.

Neural machine translation suffers poor compositionality.

PhD in Computer Science, Zhejiang University and Westlake University

Master of Science in Artificial Intelligence, University of Edinburgh

Bachelor of Engineering in Electronic Information Engineering, Wuhan University

Area Chair: ACL 2025

Conference Reviewer: ACL, EMNLP, COLING, ACL ARR, IJCAI, NeurIPS.

Journal Reviewer: TMLR, JAIR, TACL, TASLP, TBD, TALLIP.

Outstanding Student Scholarship (Silver medal, Tencent Rhino-Bird Elite Program, 2024).

National Scholarship (Ministry of Education, 2023).

Dean's Medal (Westlake University, 2023).

Website's code is from Jon Barron.