Cowork-Gym: Synthesizing Interactive Work Environments at Scale
Synthesizes 64.9K interactive work environments through a build–execute–revise loop and stateful user simulation for agent post-training.通过构建、执行与修订闭环及有状态用户模拟，合成 6.49 万个交互式工作环境，用于智能体后训练。
About me关于我
I am Xinji Mai (Shinji Mai), an Algorithm Engineer at Tongyi Lab, Alibaba Group, working on agentic reinforcement learning. My research focuses on self-evolving agents, reasoning, and scaling laws for agentic RL.我是麦新纪（Shinji Mai），目前担任阿里巴巴集团通义实验室算法工程师，从事智能体强化学习（Agentic RL）研究，关注自进化智能体、推理能力，以及智能体强化学习的规模化规律。
I received my master’s degree in Computer Technology from Fudan University and my bachelor’s degree in Microelectronics Science and Engineering from Sun Yat-sen University.我硕士毕业于复旦大学计算机技术专业，本科毕业于中山大学微电子科学与工程专业。
Research interests研究方向Agentic RLSelf-evolving Agents自进化智能体Reasoning & Scaling推理与规模化
News最新动态
- New preprint: Cowork-Gym, synthesizing 64.9K interactive work environments for agent post-training. Dataset ↗发布预印本 Cowork-Gym，合成 6.49 万个交互式工作环境用于智能体后训练。数据集 ↗
- Joined Tongyi Lab, Alibaba Group as an Algorithm Engineer.加入阿里巴巴集团通义实验室，担任算法工程师至今。
- Open-sourced the QwenPaw-Flash model family.开源模型系列 QwenPaw-Flash。
- Released CuES, a curiosity-driven synthesis framework for Agentic RL.发布 CuES，面向智能体强化学习的好奇心驱动数据合成框架。
Earlier updates更多动态
- Open-sourced AgentEvolver, an efficient self-evolving agent system.开源 AgentEvolver，高效的自进化智能体系统。
- Hi-EF accepted to AAAI 2026.Hi-EF 被 AAAI 2026 接收。
- New preprint: When Small Models Team Up, on automated intellectual-property audits.发布预印本 When Small Models Team Up，研究知识产权自动审核。
- Agentic RL Scaling Law accepted to NeurIPS 2025.Agentic RL Scaling Law 被 NeurIPS 2025 接收。
- OUS accepted to AAAI 2025.OUS 被 AAAI 2025 接收。
- Contributed multimodal-model training to Fudan's humanoid robot Guanghua-1, unveiled at WAIC 2024.参与复旦大学人形机器人光华一号的多模态大模型训练，项目在 WAIC 2024 亮相。
- Contributed to CMLM-ZhongJing, an open-source language model for Traditional Chinese Medicine.参与开源中医大语言模型仲景 CMLM-ZhongJing。
- All Rivers Run into the Sea accepted to ACM MM 2024.All Rivers Run into the Sea 被 ACM MM 2024 接收。
- Started my master’s studies in Computer Technology at Fudan University.进入复旦大学攻读计算机技术硕士学位。
- Started my undergraduate studies in Microelectronics Science and Engineering at Sun Yat-sen University.进入中山大学微电子科学与工程专业学习。
Selected publications精选论文
All publications全部论文 ↗
CuES: A Curiosity-driven and Environment-grounded Synthesis Framework for Agentic RL
Curiosity-guided exploration grounds synthetic tasks in executable environments for agentic reinforcement learning.以好奇心引导环境探索，在可执行的环境中合成智能体强化学习任务。
AgentEvolver: Towards Efficient Self-Evolving Agent System
Combines self-questioning, self-navigating, and self-attributing to help agents learn from their own experience.结合自主提问、自主探索与自主归因，让智能体从自身经验中持续学习。
Agentic RL Scaling Law: Spontaneous Code Execution for Mathematical Problem Solving
Studies how reinforcement learning teaches models to use code for mathematical reasoning, and how this ability scales.研究强化学习如何让模型自主调用代码进行数学推理，以及这种能力的规模化规律。
Hi-EF: Benchmarking Emotion Forecasting in Human-interaction
Introduces a benchmark and multimodal framework for forecasting emotions in human interactions.构建人际互动中的情绪预测基准，并提出多模态预测框架。
OUS: Bridging Scene Context and Facial Features to Overcome the Rigid Cognitive Problem
Connects scene context with facial features to support more context-aware emotion recognition.联结场景上下文与面部特征，让情绪识别更充分地理解情境。
All Rivers Run into the Sea: Unified Modality Brain-Inspired Emotional Central Mechanism
Unifies multimodal emotion signals through a brain-inspired emotional central mechanism.通过脑启发的情感中枢机制，统一建模多模态情绪信号。
Open-source models开源模型
All models全部模型 ↗
QwenPaw-Flash
Apache 2.0Tool use, command execution, memory, and multi-step planning. Original weights and Q8_0 / Q4_K_M quantizations in five sizes.面向工具调用、命令执行、记忆管理与多步规划，提供五种尺寸的原始权重与 Q8_0 / Q4_K_M 量化版本。
The Zhongjing family of language models for Traditional Chinese Medicine, with 13B and 1.8B releases.面向中医领域的 Zhongjing 系列大语言模型，涵盖 13B 与 1.8B 版本。
Open-source projects开源项目
All projects全部项目 ↗
SELF-EVOLVING AGENTS自进化智能体
AgentEvolver
Combines self-questioning, self-navigating, and self-attributing to help agents learn from their own experience.结合自主提问、自主探索与自主归因，让智能体从自身经验中持续学习。
TOOL-INTERACTIVE REINFORCEMENT LEARNING工具交互强化学习
Agentic RL Scaling Law
Asynchronous reinforcement learning and tool interaction for spontaneous Python execution and mathematical reasoning.支持异步强化学习与工具交互，让模型自主生成、执行 Python 代码，探索数学推理能力的规模化规律。
Experience & internships工作与实习经历
2026.07 – Present至今
Tongyi Lab, Alibaba Group阿里巴巴集团 · 通义实验室
Algorithm Engineer · Shanghai算法工程师 · 上海
Agentic reinforcement learning, self-evolving agents, and reasoning.智能体强化学习、自进化智能体与推理能力研究。
2025.05 – 2026.07
Tongyi Lab, Alibaba Group阿里巴巴集团 · 通义实验室
Reinforcement Learning Algorithm Intern · Shanghai强化学习算法实习生 · 上海
Self-evolving agent training, synthetic interaction data, and reward modeling; contributed to AgentEvolver.自进化智能体训练、交互数据合成与奖励建模，参与 AgentEvolver 开源项目。
2024.12 – 2025.05
Xiaohongshu · AGI Group小红书 · AGI 组
Large Language Model Algorithm Intern · Shanghai大模型算法实习生 · 上海
Mathematical reasoning data synthesis and reinforcement learning with code execution, including Agentic RL Scaling Law.数学推理数据合成与代码执行强化学习，参与 Agentic RL Scaling Law 研究。
2024.07 – 2024.12
ByteDance · TikTok字节跳动 · TikTok
Multimodal LLM Algorithm Intern · Shanghai多模态大模型算法实习生 · 上海
Multimodal models and object detection for intellectual-property audits and e-commerce platform governance.面向电商平台治理的知识产权审核、多模态模型训练与目标检测。
Education教育经历
Fudan University复旦大学
Master’s · Computer Technology硕士 · 计算机技术
2023.09 – 2026.06
Sun Yat-sen University中山大学
Bachelor’s · Microelectronics Science and Engineering本科 · 微电子科学与工程
2019.09 – 2023.06
Terramai
A little pixel world, a curious little agent.一个像素世界，一位充满好奇心的探索者。