Ph.D. student in Computer Science, Fudan University
I work on experience-guided structured reasoning for intelligent agents. My research explores how large language models can learn from prior trajectories, organize reusable reasoning skills, and make more reliable decisions in complex interactive environments.
My work sits at the intersection of reasoning, agents, and decision-making systems.
I am interested in building intelligent agents that can go beyond one-shot prompting. A central theme of my research is to transform implicit trial-and-error behavior into explicit structures such as guidelines, skills, memories, and verification procedures. Together with reliable agent harnesses, these structures make agent reasoning and action more reusable, controllable, and robust.
My recent projects study structured reasoning for LLMs, experience-induced memory for retrieval-augmented reasoning, agent harnesses, skill-oriented deployment frameworks, and decision-oriented environments such as financial markets and multi-agent strategy games.
I began my research with generative modeling, including GAN-based image synthesis and pose-guided person image generation. This early work gave me a foundation in representation learning, controllable generation, and empirical evaluation, and gradually led me toward broader questions about how intelligent systems reason, plan, and adapt.
During my Ph.D. at Fudan University, my research has shifted toward structured reasoning and LLM-based agents. I study how prior experience can be transformed into reusable guidelines, memories, and skills, and how agent harnesses can support reliable execution across reasoning, tool-use, deployment, and decision-making tasks.
I have also gained industry research experience through internships at Tencent and Microsoft, where I worked on generation quality improvement, data optimization, and prompt-based learning in real-world AI systems. I used to be a co-founder of AI2Apps, where I designed and built deployable LLM-based agent systems that connect research prototypes with practical workflows and deployment pipelines.
My research is organized around three connected directions: learning reusable reasoning experience and skills, building agentic systems with reliable harnesses, and studying decision-making agents in dynamic environments. The paper list below follows the same structure.
This direction studies how prior trajectories can be transformed into explicit guidelines, reusable skills, memories, and refinement procedures. The goal is to make LLM reasoning and agent actions more reusable, controllable, and robust across tasks.
This direction focuses on the execution layer of intelligent agents, including agent harness design, tool-use workflows, deployment pipelines, and retrieval-augmented context management. I am interested in how a harness can make agent behavior more observable, steerable, and reliable in real systems.
This direction investigates decision-making agents in dynamic environments, especially financial reasoning, risk-aware temporal analysis, and ecological market games for multi-agent strategy evolution.
Selected publications and preprints, grouped according to the three research directions above.
† Co-first authors. Public links are provided when an arXiv or TechRxiv version is available.