I am an incoming Ph.D. student at the Halıcıoğlu Data Science Institute (HDSI), UC San Diego (Fall 2026), where I am fortunate to be advised by Prof. Biwei Huang and mentored by Kun Zhou. Before that, I received my M.S. in Data Science from UC San Diego in March 2026, and my B.Sc. in Statistics (First Class Honors) from The Chinese University of Hong Kong in 2024.
My research aims to make long-horizon agentic tasks controllable and reliable through a causal lens, spanning GUI agents, long-horizon planning, and multi-agent reinforcement learning. My work follows one thread:
- Memory — equipping agents with scalable continuous memory, so that past experience is compactly reusable across unfamiliar interfaces and long horizons (CoMEM, NeurIPS 2025; CoMEM-Agent; HyMEM, ACL 2026 Findings).
- Planning — identifying planning as the dominant factor behind long-horizon performance, and efficiently improving it by training only the planner with reinforcement learning in an unbalanced multi-agent framework (Planner Matters!).
- Causal structure — harnessing agents with a unified causal structure that maintains compact, verifiable task progress, enabling evidence-based completion, failure attribution, and recovery (StructAgent).
Going forward, I aim to tackle the central challenge of long-horizon planning for agents: through verifiable state management and self-evolving memory and tools, I want agents to truly cross the sim-to-real gap — moving beyond benchmarks to deliver real productivity in the wild.
I am always open to collaborations and happy to chat about agents, causality, and everything in between — feel free to reach out!
🔥 News
- 2026.07: 📄 New preprint: StructAgent — harnessing long-horizon digital agents with unified causal structure. [Project page]
- 2026.07: 🎉🎉 I will join UC San Diego HDSI as a Ph.D. student in Fall 2026, advised by Prof. Biwei Huang!
- 2026.05: 🎉🎉 HyMEM is accepted to ACL 2026 Findings!
- 2026.05: 📄 New preprint: Planner Matters! — planning is the dominant factor in long-horizon tasks; RL-training only the planner is enough.
- 2026.03: 🎓 I graduated with an M.S. in Data Science from UC San Diego.
- 2025.10: 📄 New preprint: CoMEM-Agent — continuous memory + an auto-scaling data flywheel (100k+ trajectories) for GUI agents.
- 2025.09: 🎉🎉 CoMEM is accepted at NeurIPS 2025!
📝 Publications
(⭐️ denotes first author)
StructAgent: Harness Long-horizon Digital Agents with Unified Causal Structure
Wenyi Wu⭐️, Sibo Zhu, Kun Zhou, Aayush Salvi, Zixuan Song, Biwei Huang
arXiv preprint arXiv:2607.11388, 2026
Paper / Project Page / Code
- A unified causal structure for long-horizon digital agents: a compact, verifiable task-progress state coupled with structured workflows, enabling checkpointing, evidence-based completion, and failure recovery. Improves multiple backbones on computer-use benchmarks and generalizes to Minecraft.
Wenyi Wu⭐️, Sibo Zhu, Kun Zhou, Biwei Huang
arXiv preprint arXiv:2605.02168, 2026 (under review)
- Decomposes long-horizon automation into planner / actor / memory-manager roles and shows planning is the dominant factor in task performance — so we RL-train only the planner, yielding efficient gains across web navigation, OS control, and tool use.
Hybrid Self-evolving Structured Memory for GUI Agents
Sibo Zhu, Wenyi Wu, Kun Zhou, Stephen Wang, Biwei Huang
Findings of the Association for Computational Linguistics: ACL 2026
- HyMEM: a graph-based, self-evolving memory that couples symbolic nodes with continuous trajectory embeddings and supports multi-hop retrieval, letting small open-source backbones match or exceed closed-source models (e.g., Gemini 2.5 Pro, GPT-4o) on computer-use tasks.
Auto-scaling Continuous Memory for GUI Agent
Wenyi Wu⭐️, Kun Zhou, Ruoxin Yuan, Vivian Yu, Stephen Wang, Zhiting Hu, Biwei Huang
arXiv preprint arXiv:2510.09038, 2025 (under review)
- Encodes each GUI trajectory into fixed-length continuous embeddings (the VLM as its own encoder), sharply cutting context cost while preserving fine-grained visual cues; an auto-scaling data flywheel collects 100k+ trajectories, bringing Qwen2.5-VL-7B to the level of closed-source SOTA agents.
Towards General Continuous Memory for Vision-Language Models
Wenyi Wu⭐️, Zixuan Song, Kun Zhou, Yifei Shao, Zhiting Hu, Biwei Huang
Advances in Neural Information Processing Systems (NeurIPS), 2025
- CoMEM represents multimodal and multilingual knowledge as a compact set of dense embeddings instead of long token sequences — the key insight being that a VLM can serve as its own continuous memory encoder — with data- and parameter-efficient fine-tuning improving complex multimodal reasoning.
Causal-Copilot: An Autonomous Causal Analysis Agent
Xinyue Wang, Kun Zhou, Wenyi Wu, Har Simrat Singh, Fang Nan, Songyao Jin, Aryan Philip, Saloni Patnaik, Hou Zhu, Shivam Singh, Parjanya Prashant, Qian Shen, Biwei Huang
arXiv preprint arXiv:2504.13263, 2025
- An autonomous LLM agent that operationalizes expert-level causal analysis end-to-end — causal discovery, inference, algorithm selection, hyperparameter tuning, and interpretation — integrating 20+ state-of-the-art causal methods behind a natural-language interface.
📖 Educations
- 2026.09 – (expected 2030), Ph.D., Halıcıoğlu Data Science Institute, UC San Diego. Advisor: Prof. Biwei Huang.
- 2024.09 – 2026.03, M.S. in Data Science, UC San Diego.
- 2020.09 – 2024.07, B.Sc. in Statistics (First Class Honors), The Chinese University of Hong Kong.
💻 Internships
- Summer 2024, Tencent Holdings Limited, Shenzhen — Data Scientist Intern. Responsible for User Growth data science work in the IEG group; conducted qualitative and quantitative data analysis to investigate differentiated strategies among users.
- Summer 2023, ByteDance Technology Limited, Beijing — Data Scientist Intern. Conducted causal inference and machine learning modeling with statistical methods to analyze video and live-stream data and improve business monetization.
- Spring 2023, HSBC Insurance (Asia) Limited, Hong Kong — Data Analyst Intern.
🐱 Life
Beyond research, I am a devoted cat lover 🐈, and baking 🍰 and crocheting 🧶 are my favorite me time.
I am deeply grateful for my happy family, my beloved husband, and my dear friends — their love, support, and companionship have filled my life in California with beautiful memories ☀️.
