Bio

Xiyuan Yang (杨希渊) is now an undergraduate (sophomore) in School of Artificial Intelligence, Shanghai Jiao Tong University (SJTU-SAI). During his freshman year of study, he ranked FIRST of 62 in grades and received honors including the National Scholarship. His research interests focus on Digital Agents with broad capability boundaries (Agentic Scaling Law) and Collaborative Multi-agent Intelligence. He joined the research group MAGIC in SJTU, under the supervision of Prof. Siheng Chen.

Education

School of Artificial Intelligence, Shanghai Jiao Tong University.
GPA: 4.14/4.3 (Ranked 1 out of 62)
Score: 94.0/100
Scholarships: National Scholarship (first 3%), Zhiyuan Honor Scholarships (first 5%)
High Graded Courses:

• Comprehensive Programming Practice: 100/100
• Linear Algebra (Honor): 98/100
• Fundamental of Programming (Honor): 98/100

Research

My research interests focus on exploring the Agentic Scaling Law for large language models in the Agent era, to expand the capability boundaries of language model agents in areas such as tool calling, reasoning generalization, and long-term memory. I also investigate multi-agent collaboration workflows and information optimization within collective intelligence. I believe the future world will be a world of agents.

A list of topics that I am interested in at the moment are:

• Agentic Tool Calling and Agentic Memory Enhancement
• Multi-agent Systems and Multi-Agent Collaborations.
• Domain-augmented Agents. (Software Coding Agents, etc.)

Publications

AppCopilot: Toward General, Accurate, Long-Horizon, and Efficient Mobile Agent

Jingru Fan, Yufan Dang, Jingyao Wu, Huatao Li, Runde Yang, Xiyuan Yang, Yuheng Wang, Chen Qian

|

We introduce AppCopilot, a multimodal, multi-agent mobile agent designed for seamless cross-app operation. It implements a complete end-to-end pipeline encompassing data collection, model training, fine-tuning, efficient inference, and deployment across PC and mobile platforms. At the model level, it integrates multimodal foundation models with robust bilingual (Chinese-English) support. The reasoning and control layer employs chain-of-thought reasoning, hierarchical task decomposition, and multi-agent collaboration. The execution layer supports experiential adaptation, voice commands, function calling, cross-app and cross-device orchestration, and full mobile app compatibility. This work provides a concrete roadmap and actionable reference architecture for building general-purpose mobile agents.

InfoMosaic-Bench: Evaluating Multi-Source Information Seeking in Tool-Augmented Agents

Yaxin Du, Yuanshuo Zhang, Xiyuan Yang, Yifan Zhou, Cheng Wang, Gongyi Zou, Xianghe Pang, Wenhao Wang, Menglan Chen, Shuo Tang, Zhiyu Li, Feiyu Xiong, Siheng Chen

| | |

we introduce InfoMosaic-Bench, the first benchmark dedicated to multi-source information seeking in tool-augmented agents. Covering 6 representative domains (medicine, finance, maps, video, web, and multi-domain integration), InfoMosaic-Bench requires agents to combine general-purpose search with domain-specific tools. Tasks are synthesized with InfoMosaic-Flow, a scalable pipeline that grounds task conditions in verified tool outputs, enforces cross-source dependencies, and filters out shortcut cases solvable by trivial lookup. This design guarantees both reliability and non-triviality.

Project

Skills

Programming Languages

Tools

Python Modules

ML & DL

• Deep learning architectures
• Model training techniques
• Reinforcement learning

LLM & Agents

• Large Language Models
• Post training techniques
• Multi-agent frameworks

Languages

• Chinese (native)
• English (CET-6: 584)

Technical Blogs

|

Active Github committer with 30+ open-source repositories and 1500+ commits.

Maintainer of Xiyuan Yang's Technical Blog. I regularly publish technical content focusing on computer science and AI. To date, I have authored over 120 articles with a cumulative word count exceeding 450,000 words.

Technical Projects

More will be added here in the future. You can see Tool Zoo for more info.

AlphaBuild: Generating Formulaic Alphas on a Wider Range of Stock Data

Leading AlphaBuild as the final project of course SJTU-AI1803, an RL-based methodology using PPO to optimize factor mining process based on AlphaGen, achieving superior backtest performance by leveraging smaller, parallel factor pools and PCA for final factor selection.

ProbeCode: AI coding Agent with MCP Framework

ProbeCode: AI coding agent integrating static code inspection with a ReAct framework to understand and memorize long-context code. It operates and comprehends code at the project level and tackles the core challenge of long and complex codebases exceeding standard LLM context windows.

Awards

• 2025-10: Challenge Cup Defender (揭榜挂帅擂台赛 人工智能赛道 擂主 特等奖第一名)
• 2025-09: National Scholarship (first 3%)
• 2025-01: Honorable Mention in MCM 2025
• 2024-10: Zhiyuan Honor Scholarships (first 5%)