Zejun (Warrior) Xu

zejunx@student.must.edu.mo | warriorxu0302.github.io | github.com/WarriorXu0302

Research Interests: GUI / Browser Agents, Vision-Language Models, Reinforcement Learning

Education

Macau University of Science and Technology (MUST)Sept. 2025 – Present

M.S. in Intelligent TechnologyMacau, China

Jilin International Studies UniversitySept. 2021 – June 2025

B.S. in Artificial IntelligenceJilin, China

Selected Awards: 3 National Awards in Ministry-of-Education Category A competitions.

Publications

Zejun Xu*, Taiyi Chen*, Jin Li*, Yongtong Gu*, Qi Cheng, Aixuan Lv, Shuai Zhu, Pengfei Zhu, Kaichen Yang, Boyu Sun, Yixian Yang, Mulong Xie, Xin Liu, Dagang Li, Xiaoteng Ma, Hongru Wang.
CAP: A Scalable Benchmark for Evaluating Cross-Site Browser Agents with Complex Actions and Perception.
Third Conference on Language Modeling (COLM 2026). [OpenReview] [Code] [Project Page]
420 cross-site tasks over 108 real-world websites, evaluated by a verifiable agent-as-a-judge framework. Earlier version accepted at the Lifelong Agent Workshop @ ICLR 2026. *Equal contribution.

Research & Work Experience

Macaron AI (Mind Lab)Jan. 2026 – Mar. 2026

Research Intern (Part-time, Hybrid)

Model Distillation: Studied On-Policy Distillation algorithms for more efficient training and inference.

Tsinghua University (TsinghuaNLP) & OpenBMBMay 2025 – Apr. 2026

Research Assistant (Part-time, Hybrid) — Advisor: Prof. Zhiyuan LiuBeijing, China

Model Training: Conducted SFT, LoRA, and full-parameter fine-tuning of MiniCPM models.
Reinforcement Learning: Applied PPO & GRPO to improve agent performance in RL environments.
Agent Development: Built backend systems for AgentCPM and Tsinghua QingXiaoDa; developed MCP tools (stdio) for LLM tool use.

Fellou AIMay 2025 – Apr. 2026

Algorithm & Benchmark Engineer (Part-time, Hybrid)Beijing, China

CAP-Bench: Co-led construction of CAP-Bench (420 tasks, 108 real-world websites) and its verifiable agent-as-a-judge evaluation framework to evaluate and optimize Fellou's GUI capabilities; first-author paper accepted at COLM 2026.
Model Optimization: Fine-tuned, quantized, and distilled open-source models (Qwen series).

Institute of Computing Technology, CAS (Food Computing Group)Sept. 2024 – Nov. 2024

Deep Learning Algorithm Intern — Advisor: Prof. Shuqiang JiangSuzhou, China

LLM Fine-tuning: Conducted SFT for the “Shitian” large model and built a 100K+-entry training dataset.
RAG: Implemented a RAG pipeline to improve factual quality of model outputs.

iFlytek Co., Ltd.Apr. 2024 – June 2024

Assistant CV Engineer (Smart Robot Group)Hefei, China

Optimized 3D point-cloud perception and human-pose models; built the “Point-and-Sweep” finger-pointing feature and automated test suites for cleaning robots.

Open-Source Projects

AgentDesk — Enterprise Multi-User Agent PlatformMay 2026 – Present

Creator & primary author (83% of commits)github.com/WarriorXu0302/agentdesk-platform

Designed and built an open-source, business-agnostic multi-user agent platform in TypeScript (48K+ LOC, 855 tests, 45 architecture decision records): multi-channel ingress, per-user session isolation, and frontdesk-to-worker agent delegation.
Engineered an unforgeable end-user identity trust chain across agent hops — HMAC-signed gateway requests, host-side identity cross-validation, fail-closed security defaults, and a central audit table for every backend call.
Implemented per-session containerized execution with cgroup resource caps, user-level RBAC with multi-tenant organization isolation, and Prometheus + OpenTelemetry observability.

SlideGen — AI-Powered Presentation EditorMar. 2026

Sole author, MIT-licensedgithub.com/WarriorXu0302/SlideGen

Built an LLM-driven presentation tool with natural-language slide generation, in-browser visual editing, and PPTX import/export.

Technical Skills

Agent Engineering: Model Context Protocol (MCP), A2UI, Playwright, GUI Agent, RAG.
LLM/VLM Training: SFT, On-Policy Distillation, RL (PPO, GRPO), LoRA/QLoRA, CoT.
Frameworks & Tools: Python, SQL, PyTorch, Hugging Face Transformers, Git, Docker, Linux, WandB.

Other Publication: Xinzhu Liu, Shiyu Zhou, Yuhe Wang, Zejun Xu, et al. Research on the Influence Mechanism of Tourism Behavior Intention of Chinese and Russian Tourists in Northeast China Based on PLS. Areal Research and Development, 2025, 44(3): 89–95. (CSSCI)