Zihao Zhou

Zihao Zhou 周梓浩
Google Scholar / Github / Twitter
[email protected]

Hi! I am a CS PhD student at University of Liverpool and Xi’an Jiaotong-liverpool University, beginning from 2022 Fall. Recently, I am working closely with Wenda Li and Meng Fan on informal mathematical proving.

My research interests mainly lie in:

• Large Language Models Reasoning: Advancing the frontier mathematical capabilities of AI systems.
• Scalable Reasoning: Exploring scalable methods to enhance the complex reasoning capabilities of foundation models.
• Real-World Applications : Leveraging LLMs to tackle open-ended real-world challenges.

Publications

• [Preprint] Numina-Lean-Agent: An Open and General Agentic Reasoning System for Formal Mathematics. [Demo] [paper] [code]
  • Junqi Liu*, Zihao Zhou*, Zekai Zhu*, Marco Dos Santos, Weikun He, Jiawei Liu, Ran Wang, Yunzhou Xie, Junqiao Zhao, Qiufeng Wang, Lihong Zhi, Jia Li, Wenda Li
  • 🔮 We propose Numina-Lean-Agent, a general agentic reasoning system that can autonomously interact with diverse reasoning tools. Numina-Lean-Agent achieves state-of-the-art performance on Putnam 2025 (12/12) and successfully formalizes the Brascamp–Lieb theorem in collaboration with mathematicians.
    
    
• [Preprint] F1-Reasoner: Synthesizing Verifiable Reasoning Data From Formal Math Statements
  • Zihao Zhou, Wei Liu, Xinlong Fu, Kaizhu Huang, Xiaowei Huang, Meng Fan, Wenda Li, Qiufeng Wang
  • 🏁 We introduce F1-Reasoner, a framework for synthesizing high-quality verifiable reasoning data from formal mathematical statements. Both F1-Reasoner and its Mix version outperform baselines that rely on either synthetic data from artificial environments or human data.
    
    
• [NeurIPS 2025] Can MLLMs Absorb Math Reasoning Abilities from LLMs as Free Lunch? [paper]
  • Yijie Hu*, Zihao Zhou*, Kaizhu Huang, Xiaowei Huang, Qiufeng Wang
  • 👁️ We propose a cross-modal model merging approach IP-Merging, which can transform math reasoning ability from LLMs to MLLMs. Without training, we can achieve 4.8% improvement on MathVista by merging Qwen2-vl-instruct and Qwen2-Math-Base. In addition, we conduct detailed analyses in parameters selection and projection.
    
    
• [Preprint] GeoSDF: Plane Geometry Diagram Synthesis via Signed Distance Field. [paper]
  • Chengrui Zhang*, Maizhen Ning*, Tianyi Liu, Zihao Zhou, Jie Sun, Kaizhu Huang, Qiufeng Wang
  • 📏 GeoSDF is an SDF-based tool that can synthesize plane geometry diagram and verifiable signal(like length, angle) by only offering formal relation description. It achieves 95.9% on GeoQA by synthesize-then-measure, highlighting its potential as a plane geometry environment.
    
    
• [ICLR 2025] Is Your Model Really A Good Math Reasoner? Evaluating Mathematical Reasoning with Checklist. [website] [paper] [code]
  • Zihao Zhou*, Shudong Liu*, Maizhen Ning, Wei Liu, Jingdong Wang, Derek F. Wong, Xiaowei Huang, Qiufeng Wang, Kaizhu Huang
  • 🔍 MathCheck includes multiple mathematical reasoning tasks and robustness test types to facilitate a comprehensive evaluation of both mathematical reasoning ability and behavior testing.
    
    
• [AAAI 2025] GNS: Solving Plane Geometry Problems by Neural-Symbolic Reasoning with Multi-Modal LLMs. [paper][code]
  • Maizhen Ning*, Zihao Zhou*, Qiufeng Wang, Xiaowei Huang, Kaizhu Huang
  • GNS is a neural-symbolic reasoning system simultaneously capable of multiple geometry reasoning tasks beyond solving. Training: multi-tasks SFT. Inference: combining knowledge prediction, geometry image parsing and tool calling.
  • Selected as Oral Presentation(Top 5%) 🎉.
    
    
• [AAAI 2024] MathAttack: Attacking Large Language Models Towards Math Solving Ability. [paper] [code]
  • Zihao Zhou, Qiufeng Wang, Mingyu Jin, Jie Yao, Jianan Ye, Wei Liu, Wei Wang, Xiaowei Huang, Kaizhu Huang
  • 😈 Examining the robustness of LLMs in math reasoning ability by textual attack algorithm.
    
    
• [ACL 2023 Findings] Learning by Analogy: Diverse Questions Generation in Math Word Problem. [paper] [code]
  • Zihao Zhou*, Maizhen Ning*, Qiufeng Wang, Jie Yao, Wei Wang, Xiaowei Huang and Kaizhu Huang
  • ✍️ A framework to generate diverse questions with labels for a given math word problems.
    
    
• [NLPCC 2023] Solving Math Word Problem with Problem Type Classification. [paper][code]
  • Jie Yao*, Zihao Zhou*, Qiufeng Wang
  • A problem type classifier to combine the abilities of LLM solver and traditional math solver.
  • 🎯 Ranked 2nd in NLPCC 2023 Share Task 3.
    
    
• [ICASSP 2021] Knowledge-Based Chat Detection with False Mention Discrimination. [paper]
  • Wei Liu, Peijie Huang, Dongzhu Liang, Zihao Zhou
  • A two-stage pipeline for chat detection 🗣️🤖.

Experience

• Tsinghua University, Visiting Student at KEG Lab, 2023.09 - 2024.03
• University of Liverpool, PhD Student, 2022.09 - Present
• South China Agricultural University, Bachelor, 2018.09 - 2022.06

Services

• Conference Reviewer: ICLR 24/25/26, ICML 24/25, NeurIPS 24, ACL ARR 23/24
• Journal Reviewer: TASLP, Neural Networks