Jiayi Kuang
况佳杙

Master's Student in Control Science and Engineering

Sun Yat-sen University, Guangzhou, China

Research Interests

Large Language Models • Multimodal Learning • Mathematical Reasoning • AI Agents

I am currently a Master's student at Sun Yat-sen University and a research intern at Tencent Youtu Lab. My research focuses on mathematical reasoning in large language models, native multimodal large models, and LLM pre-training. I am enthusiastic about academic collaborations and welcome opportunities to work together. Feel free to reach out!

📧 Email: kuangjy6@mail2.sysu.edu.cn

📱 Phone: (+86) 187-2568-1142

Google Scholar GitHub CV
Jiayi Kuang

Education

2024.9 - Present

Master of Engineering

Control Science and Engineering

Sun Yat-sen University, Shenzhen, China

Research Focus: Large Language Model Reasoning, Multimodal Large Language Models

Advisor: Prof. Ying Shen

2020.9 - 2024.7

Bachelor of Science

Electronic Information Science and Technology

Sun Yat-sen University, Shenzhen, China

Research Focus: Multimodal Computing, Visual Question Answering

Advisor/Collaborator: Prof. Yulan Guo

Publications

12 papers published/accepted in top-tier international conferences and journals (7 as first/co-first author)

ICLR 2026

Process-level Trajectory Evaluation for Environment Configuration in Software Engineering Agents

Jiayi Kuang, et al. (First author)

International Conference on Learning Representations (ICLR 2026)

Proposed process-level trajectory evaluation methods for software engineering agents, focusing on environment configuration optimization to improve agent performance in code generation and debugging tasks.

NeurIPS 2025

Atomic Thinking of LLMs: Decoupling and Exploring Mathematical Reasoning Abilities

Jiayi Kuang, et al. (First author)

Conference on Neural Information Processing Systems (NeurIPS 2025)

Decoupled mathematical reasoning atomic abilities of LLMs from domain knowledge and logical thinking perspectives, exploring interactions between different atomic abilities including synergistic and antagonistic effects.

ACL 2025

Browsing Like Human: A Multimodal Web Agent with Experiential Fast-and-Slow Thinking

Haohao Luo, Jiayi Kuang, et al. (Second author)

Annual Meeting of the Association for Computational Linguistics (ACL 2025, CCF-A, Long Paper)

Designed a fast-slow thinking planning mode that mimics human cognition, incorporating reflection on past successes and failures to enhance web search agent performance.

ACL 2025

Express What You See: Can Multimodal LLMs Decode Visual Ciphers with Intuitive Semiosis Comprehension

Jiayi Kuang, et al. (First author)

Annual Meeting of the Association for Computational Linguistics (ACL 2025, CCF-A, Long Paper)

Addressed the lack of visual sensitivity and visual intuition in existing benchmarks by proposing a benchmark that mimics human visual intuition semiotics.

ICML 2025

One Example Shown, Many Concepts Known! Counterexample-Driven Conceptual Reasoning in Mathematical LLMs

Yinghui Li, Jiayi Kuang*, et al. (Co-first author)

International Conference on Machine Learning (ICML 2025)

Investigated the role of counterexample reasoning in mathematical reasoning for LLMs. Trained a 7B model with 1,025 counterexamples, achieving significant improvements across multiple benchmarks, even surpassing 72B models.

ICLR 2025

Refine Knowledge of Large Language Models via Adaptive Contrastive Learning

Yinghui Li, Haojing Huang, Jiayi Kuang*, et al. (Co-first author)

International Conference on Learning Representations (ICLR 2025)

Explored the honesty hallucination problem in large language model generation. Utilized dynamic adaptive contrastive learning to consolidate known knowledge and forget unknown knowledge, achieving hallucination mitigation.

ACM CS

Natural Language Understanding and Inference with MLLM in Visual Question Answering: A Survey

Jiayi Kuang, et al. (First author)

ACM Computing Surveys 2025 (JCR Q1, IF: 28.0)

Deconstructed visual question answering into natural language (including visual and textual languages) understanding and reasoning, systematically reviewing the evolution from traditional deep learning methods to multimodal large language models.

Internship Experience

Research Intern at Tencent Youtu Lab

July 2025 - Present

"Project up Program" (青云计划) Research Intern

Participating in the Youtu-LLM project, focusing on unlocking the native agentic potential for lightweight large language models. Working on developing efficient agent frameworks and optimization strategies to enhance the autonomous capabilities of compact LLMs for real-world applications.

Key Project: Youtu-LLM: Unlocking the Native Agentic Potential for Lightweight Large Language Models

Technical Skills

Programming & Tools

  • Python, MATLAB
  • PyTorch, TensorFlow
  • NumPy, Pandas, NLTK
  • Git, Linux

Research Expertise

  • Large Language Model Training & Fine-tuning
  • AI Agent Frameworks
  • Retrieval-Augmented Generation (RAG)
  • Multimodal Learning

Core Competencies

  • Natural Language Processing
  • Mathematical Reasoning
  • Computer Vision
  • Data Structures & Algorithms

Contact

Email

kuangjy6@mail2.sysu.edu.cn

Phone

(+86) 187-2568-1142

Location

Guangzhou, Guangdong, China