⭐
⭐
profile photo

Chenxin Li | 李宸鑫

Hi! I'm Chenxin Li, a Ph.D. student at The Chinese University of Hong Kong (CUHK). I earned my master's & bachelor's degrees from Xiamen University, along with a second major in Economics.

I work on Long-Horizon Code/CLI Agents, especially for:

(i) Code/CLI Agents for Long-Horizon Software Workflows: CLI Agent Mid-training / Harness Stability (Seed2.0); Frontend Coding / Code-to-Chart (Seed1.8); CLI/MCP Tool-Use (UI-TARS-2); Claw-Eval (Live); GameCraft-Bench; PhoneBuddy Series; IR3D-Bench; PSAgent.
(ii) Auto AI Self-improvement via Data, Rollouts, and Verifiers: Auto R&D Agent ("Seed-for-Seed" prototype in Seed2.1); On-Policy Data Evolution.

I did intern@ByteDance Seed, Tencent Hunyuan, Tencent AI Lab, etc; won PaperDigest Most Influential Paper@AAAI'25; explored agent workflows via 🛠️ OpenAgentLabs and expressed insights via 📝 Blogs. Always open to discussion and collaboration and please reach out via email (chenxinli@link.cuhk.edu.hk) or wechat (jasonchenxinli).

LinkedIn | Google Scholar Scholar | GitHub | X | Blogs | OpenAgentLabs

🚀 Selected Works [Google Scholar]
* Equal contribution, † Project Leader, ‡ Corresponding author
Seed2.0
Seed2.0: Towards Intelligence Frontier for Real-World Complexity
ByteDance Seed Team (including Chenxin Li)

[Project] [Model Card]

Contributed to CLI agent capacity of generalized performance under different harnesses.
Seed-1.8
Seed-1.8: Towards Generalized Real-World Agency
ByteDance Seed Team (including Chenxin Li)

[Project] [Model Card]

Contributed to frontend coding and code-to-chart capacity.
UI-TARS-2
UI-TARS-2: Advancing GUI Agent with Multi-Turn Reinforcement Learning
ByteDance Seed Team (including Chenxin Li)

[Project] [Report]

Contributed to CLI and MCP tool-use capacity.
Claw-Eval-Live Overview
Claw-Eval (Live): A Live Agent Benchmark for Evolving Real-World Workflows
Chenxin Li†, Zhengyang Tang, Huangxin Lin, Yunlong Lin, Shijue Huang, Shengyuan Liu, Bowen Ye, Rang Li, Lei Li, Benyou Wang, Yixuan Yuan

[Project] [Paper] [Code] [AI 生成未来 | PaperAgent]

A live workflow-agent benchmark with refreshable demand signals and verifiable execution traces; 105 tasks across 22 categories, 13 frontier models, top model passes only 66.7%.
IRBlender-Bench Framework
IR3D-Bench: A Blender-CLI Agent Benchmark for Inverse Rendering
Parker Liu*, Chenxin Li*, Zhengxin Li, Yipeng Wu, Wuyang Li, Zhiqin Yang, Zhenyuan Zhang, Yunlong Lin, Sirui Han, Brandon Y. Feng
NeurIPS 2025

[Project] [Paper] [Code]

A Blender-CLI code-agent benchmark that closes the loop from visual understanding to structured code generation, Blender execution, and environment feedback.
GameCraft-Bench Overview
GameCraft-Bench: Can Agents Build Playable Games End-to-End in a Real Game Engine?
Tongxu Luo, Rongsheng Wang, Jiaxi Bi, Chenming Xu, Zhengyang Tang, Jianlong Chen, Juhao Liang, Ke Ji, Shuqi Guo, Yuhao Du, Fan Bu, Wenyu Du, Xiaotong Zhang, Kyle Li, Shaobo Wang, Linfeng Zhang, Yuxuan Liu, Xin Lai, Chenxin Li, Yiduo Guo, Zhexin Zhang, Xinyuan Wang, Tianyi Bai, Ziniu Li, Benyou Wang

[Project] [Paper] [Code]

A game-generation benchmark for coding agents, with 140 Godot tasks and interactive verification of playable end-to-end game artifacts.
PhoneBuddy Figure 1
PhoneBuddy Series: Training Open Models for Agentic Phone Use

[Project] [Paper] [Code]

An open phone-use agent line trained with real-phone execution, scalable mock-app environments, and verifier-backed rollouts.
PhoneHarness Figure 1
HybridClaw: A Mixed-Action Orchestration Harness and Benchmark for Phone Agents across CLI, GUI, and MCP Tools

[Project] [Paper] [Code] [Dataset] [机器之心]

A mixed-action phone-agent harness and benchmark that routes across CLI, GUI, and MCP tools with trace-backed verification.
🛠️ Open Agent Labs [Org]

I build agent tools for AI-native workflow automation.

Open Codex Computer Use preview Open Codex Computer Use Background computer-use MCP for agents. Open Codex Mobile cover Open Codex Mobile Control desktop Codex from a mobile app. Open Typeless Harness Voice-coding harness for vibe coding. Agent Context Manager cover Agent Context Manager Cross-session context for coding agents.

More: LinMem: lightweight agent memory; AI-native Slides: agent-friendly slides; OpenReview Agent: OpenReview automation.

🧭 Selected Experience [LinkedIn]
  • Meta Superintelligence Labs, Incoming Intern, coding agents
  • ByteDance Seed, Intern, CLI agents, frontend coding, and CLI/MCP tool-use
  • Tencent Hunyuan, Intern, phone-use agents and live workflow-agent evaluation
  • Tencent AI Lab, Intern, Blender-CLI agents and structured visual reasoning
  • ScholaGO, Co-founder & Tech Lead, 2023-2024: K-12 education-agent platform, supported by HKSTP, HK Tech 300, and Alibaba Cloud
🎓 Professional Activities
  • Workshop Organizer: AIM-FM: Advancements In Foundation Models Towards Intelligent Agents (NeurIPS 2024)
  • Talks: VALSE Summit (Jun 2025) and DAMTP, University of Cambridge (Jul 2024)
  • Conference Reviewer: ICLR, NeurIPS, ICML, ACL, CVPR, ICCV, ECCV, EMNLP, AAAI, ACM MM
  • Journal Reviewer: Nature Machine Intelligence, PAMI, TIP, DMLR, PR, TNNLS
🌱 Hobbies Beyond Work
  • Vibe Working: I build and use agents as daily collaborators, testing how much work and life can be wired into agent workflows.
  • Reading: I read history, philosophy, and sociology to reason about people, systems, and long-term trends.
  • Stock Investment: I treat investing as real-world RL: making decisions under uncertainty, learning from feedback, and allocating capital, attention, and time.