Agent Learning & Post-training
OPD, RSI-based self-improvement, synthetic data for code agents, and multi-turn reinforcement learning.
Explore researchAbout me!
Zishuo ZHAO (赵子硕) is a Large Language Model Algorithm Engineer at Alibaba Holding, which he joined on August 3, 2026. His current work spans code agent data synthesis, multimodal agent training, and research on agent learning. Previously, he held the same role at Alibaba International from July 4, 2024 to July 3, 2026, applying agent research to customer service, voice interaction, and multimodal decision-making.
He received his master's degree in Intelligent Control from the School of Intelligent Systems Engineering at Sun Yat-sen University in 2024, and his bachelor's degree in Computational Materials from the same university in 2021. Previously, he interned at Tencent AI Lab and Alibaba International.
My research interests include LLM agents and reinforcement learning, natural language processing, and multimodal learning. I am particularly interested in learning from multi-turn interactions, aligning agents with human preferences, and improving their reasoning and tool use.
Research Directions
Multimodal Learning
Multimodal agents, cross-modal retrieval, multimodal graph learning, and vision-language applications.
Explore researchFor research discussions and collaboration, please feel free to get in touch.
News
Our work on turn-level alignment for LLM agents is available on arXiv.
Joined Alibaba Holding as a Large Language Model Algorithm Engineer.
Joined Alibaba International.
Our source-free ABSA paper appeared at LREC-COLING 2024.
Publications (Google Scholar)
Research on agent learning, language understanding, and multimodal representations.
Experience
Alibaba Holding
Large Language Model Algorithm Engineer
Code Agent Data Synthesis
Develop synthetic training data for code agents, focusing on Terminal-Bench 3.0, machine learning, and scientific tasks. Work on task-oriented data and agent interaction trajectories to support coding, experimentation, and problem solving across these domains.
Multimodal Agent Training
Train multimodal models as agents, covering reflection, OPD, tool calling, and revisiting prior observations. Work on reinforcement learning for multimodal agent tasks that combine perception, reasoning, and interaction across multiple steps.
Research on Agent Learning
Investigate Claude's reasoning trajectories through probing and reverse engineering. Explore knowledge distillation, OPD, RSI-based self-improvement, and agentic reinforcement learning, with a focus on reasoning transfer and iterative agent improvement.
Alibaba International
Large Language Model Algorithm Engineer
I worked on applying language and multimodal models to real-world customer interactions, with a focus on agent planning, tool use, preference alignment, and evaluation.
Customer Service Agents
Contributed to pre-sales and after-sales agents for AliExpress, including planner–reasoner architectures, tool calling, reward evaluation, and preference alignment with DPO and GRPO.
Multimodal Review
Contributed to image-based review workflows using multimodal models, OCR, image classification, and preference alignment.
Voice Agents
Worked on Lazada inbound customer service and ICBU outbound voice agents, covering planning, knowledge grounding, interruption handling, and order understanding.
Internships
Alibaba International
Research Intern
Explored RLHF and DPO for customer service language models, reinforcement learning frameworks, and automated model evaluation.
Tencent AI Lab
Research Intern
Worked on LLM role-playing, real-time game commentary, and game report generation. Participated in research on PPO-based methods for extending model input context.
Education
Sun Yat-sen University
Master's degree in Intelligent Control
School of Intelligent Systems Engineering · Shenzhen, China
Sun Yat-sen University
Bachelor's degree in Computational Materials
School of Materials · Guangzhou, China
Let's connect.
LLM agents · Reinforcement learning · NLP · Multimodal learning