Xiyuan Yang (杨希渊)

About Me

Xiyuan Yang (杨希渊) is now an undergraduate in School of Artificial Intelligence, Shanghai Jiao Tong University (SJTU-SAI). He ranked SECOND of 62 in grades and received honors including the National Scholarship (5 scholarships in total). He joined the research group MAGIC in SJTU, under the supervision of Prof. Siheng Chen.

My research interests lie in continuously pushing the capability boundaries of language intelligence. I focus on two key directions: scalable agentic data synthesis and training pipeline, which generates high-quality instances and agent trajectories for mid-training and post-training, with a particular emphasis on improving performance on long-horizon tasks; and recursive self-improvement (RSI) systems, which enable efficient, iterative model improvement through both training-time and inference-time approaches, with a particular emphasis on enhancing models' capabilities for automated post-training.

Topics I'm currently interested in:

Self-Evolving Agents for Autonomous LLM Post-Training (RSI)
Long-Horizon Agentic Instance Construction and Training
Agentic Harness Reinforcement Learning

Download my CV

Education

Shanghai Jiao Tong University

B.S. in Artificial Intelligence, School of Artificial Intelligence

2024 - 2028

GPA: 4.1/4.3 (Ranked 2 out of 62)

Score: 94.0/100

Scholarships: National Scholarship (first 3%), "Han Ying Ju Hua" Scholarship (15 per year), Zhiyuan Honor Scholarships (first 50%), SJTU Undergraduate Excellence Scholarship

High Graded Courses:

Comprehensive Programming Practice: 100/100
Probability and Statistics (Honor): 100/100
Algorithm Design and Analysis: 100/100
Numerical Analysis: 100/100
Linear Algebra (Honor): 98/100
Fundamentals of Programming (Honor): 98/100
Fundamentals of ML, DL and RL: 98/100

Experience

IQuest Research (IQuest Coder)

Internship, LLM Post-Training

2026.7 - Present

MAGIC Lab

Undergraduate Research Assistant

Advisor: Prof. Siheng Chen

2024.12 - Present

Publications & Tech Projects

Technical Report

IQuest-Q1: Advancing Foundation Capabilities for Agentic CLI Systems

Denoted as contributors.

Github / HuggingFace / Web

IQuest-Q1 is a Mixture-of-Experts (MoE) model developed by IQuest for agentic coding, reasoning, and multi-step tool use. It comprises approximately 320B total parameters, with an estimated 15B parameters activated per token. Built for CLI agents, it combines general-purpose, coding-agent, and CLI interaction capabilities, working through coding and general agentic tasks in the command line, from reading the codebase and running tools to checking results and recovering from errors. To develop these capabilities over medium- and long-horizon tasks, we train the model in synthetic environments using multi-harness reinforcement learning and multi-teacher on-policy distillation (MOPD).

RSI-Master: Structuring Experiments to Guide Autonomous Model Improvement

Yaxin Du*, Xiyuan Yang*, Zhifan Zhou, Yujie Ge, Cheng Wang, Jiajun Wang, Sijie Chen, Zehui Liu, Yuxin Zhang, Weicheng Gu, Julian Zhang, Zixing Lei, Siheng Chen

arXiv / Code

We introduce RSI-Master, which regularizes step-wise actions via an Experiment OS and structures exploration through Reviewer-Guided Research Orchestration over a growing research DAG of Workers and Reviewers, avoiding both hacking and strategy lock-in in autonomous model development. On PostTrainBench with Qwen3-4B-Base it averages 54.49 versus 46.53 for the strongest agent baseline at a 0.0% hacking rate, and at 35B surpasses the human-developed Instruct model on LiveCodeBench-v6 (41.21 vs. 37.36) and SciCode.

DataMaster: Data-Centric Autonomous AI Research

Yaxin Du*, Xiyuan Yang*, Zhifan Zhou, Wanxu Liu, Zixing Lei, Zimeng Chen, Fenyi Liu, Haotian Wu, Yuzhu Cai, Zexi Liu, Xinyu Zhu, WenHao Wang, Linfeng Zhang, Chen Qian, Siheng Chen

arXiv / Code / RedNote

We introduce DataMaster, a data-agent framework for task-conditioned autonomous data engineering. Given a fixed learning algorithm, DataMaster improves downstream performance by searching for external data, selecting and composing datasets, and applying cleaning or transformation. Its DataTree, shared Data Pool, and Global Memory coordinate branch exploration and reuse evidence, improving medal rate by 32.27% on MLE-Bench Lite and surpassing the instruct model on GPQA.

ICML 2026 Regular

MCP-Persona: Benchmarking LLM Agents on Real-World Personal Applications via Environment Simulation

Wenhao Wang*, Peizhi Niu*, Gongyi Zou*, Xiyuan Yang*, Jingxing Wang*, Haoting Shi, Yaxin Du, Jingyi Chai, Xianghe Pang, Shuo Tang, Yanfeng Wang, Siheng Chen

arXiv / Code

We introduce MCP-Persona, a benchmark for evaluating LLM agents on personalized MCP tools in realistic social and productivity environments. It simulates account- and database-grounded tasks across Reddit, Xiaohongshu, Lark, and Slack, exposing challenges beyond generic information seeking. Experiments with SOTA agents reveal substantial gaps in personalized tool use, highlighting MCP-Persona as a practical testbed for personal-application agents.

ICLR 2026 Poster

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

arXiv / Dataset / Code / Project

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.

Technical Report

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

arXiv / Code / Models

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.

Open Source Projects

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

Open-Source Projects

IntelliSearch V3.1: Unifying Search, Empowering Action for tool calling autonomous agents.

SAI Community: The first open-source SAIer's forum for courses, careers and future.

Technical Blog: 120+ articles on CS and AI, 450k+ words.

LLM Infra Docs: Core source code analysis of large-scale open-source LLM infrastructure.

Agent Codebase: Generalized codebase for agentic scientific research pipelines.

PaperFlow: Never use Overleaf again!

SyncFlash: Lightning-fast configuration of your server.

Course Labs

LLM Reasoning: Enhancing reasoning abilities for LLMs using reinforcement learning.

Clustering: Clustering of high-dimensional data with intrinsic low rank.

Image Scaling: 2D image scaling based on classical interpolation algorithms.

CTR-Press: Compress-then-Refine for training-free KV cache compression.

L-PHYM: Long-horizon language-driven physics-based motion control.

Data Structure: Course notes and source code for data structures.

CSAPP: Computer Systems: A Programmer's Perspective (labs and notes).

LLM: Stanford CS336 language models learning notes.

Technical Blogs

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.

GitHub Repository / Blog Website

I also maintain Frontier LLM Notes (https://xiyuanyang-code.github.io/llm/), an actively growing set of write-ups tracking the fast-moving LLM field. It covers model architectures, pretraining / midtraining / post-training recipes, concrete training techniques, core infrastructure with source-level LLM codebase analysis, and agentic LLMs.