Xixi WU 「吴茜茜」

Ph.D. candidate,
The Chinese University of Hong Kong

Email / GitHub / Google Scholar

Short Bio

I am a Ph.D. student at The Chinese University of Hong Kong, under the supervision of Prof. Hong CHENG.
Previously, I received my B.S. and M.S. in Computer Science from Fudan University in 2021 and 2024, respectively, under the supervision of Prof. Yun Xiong. I was a Research Intern at Microsoft Research Asia (Shanghai AI/ML Group), mentored by Dr. Yifei SHEN and Dr. Caihua SHAN.

I have been with the Tongyi DeepResearch team since June 2025, conducting research on Agentic RL to unlock long-horizon search intelligence for web agents. Prior to my research on LLMs, I conducted research on Data Mining, focusing on topics like community detection, graph prompt learning, recommender systems, etc.

Recent News

  • 🔥 [Sep 2025] Excited to share our team's progress: Tongyi-DeepResearch-30B-A3B achieves SOTA performance across open-source and proprietary agents on multiple benchmarks, accompanied by two papers detailing the methods behind these gains, ReSum, a minimal extension to ReAct that enables long-horizon search, and WebSailor-v2, a scalable data-synthesis and RL training pipeline.
  • [Aug 2025] Our paper MoLoRAG was accepted to EMNLP'2025. See you in Suzhou ✨
  • [July 2025] Introducing WebSailor, which achieves open-source SOTA performance on some of the most difficult browsing benchmarks. WebSailor topped the HuggingFace daily papers 🥇
  • [May 2025] Our paper LLMNodeBed was accepted to ICML'2025. See you in Vancouver again 😄
  • [Nov 2024] I was awarded NeurIPS'2024 Top Reviewer.
  • [Sep 2024] Our paper GNN4TaskPlan was accepted to NeurIPS'2024. See you in Vancouver!

Research Highlights

  • generation

    ReSum: A New Inference Paradigm for Web Agents

    We introduce ReSum, a novel paradigm that enables indefinite exploration through periodic context summarization. ReSum minimizes modifications to ReAct to avoid additional architectural complexity, ensuring plug-and-play compatibility with existing agents. We further design ReSum-GRPO to familiarize agents with this paradigm. Experiments across web agents on three challenging benchmarks show average improvements of 4.5% for ReSum compared to ReAct, with further improvements of 8.2% after ReSum-GRPO training.

  • generation

    WebSailor: Open recipe for Post-training on Web Agents

    WebSailor provides an open recipe for post-training web agents, including data synthesis, supervised fine-tuning, and reinforcement learning. Built on the Qwen3-30B-A3B model, WebSailor-V2 achieves state-of-the-art performance, scoring 35.3% on BrowseComp-en, 44.1% on BrowseComp-zh, and 30.6% on Humanity’s Last Exam (HLE).

  • generation

    Comprehensive Analysis of LLM4Graph Algorithms (ICML'2025)

    We introduce LLMNodeBed, a PyG-based testbed for LLM-based node classification algorithms. It integrates 14 datasets, supports 8 LLM-based and 8 classic methods, and covers 3 learning configurations. With LLMNodeBed, we train and evaluate over 2,700 models to analyze the effects of factors like LLM type and size, prompt, learning paradigm, and dataset homophily. Our study reveals 8 key insights, including (1) optimal settings for each algorithm category, and (2) scenarios where LLMs significantly outperform LMs.

  • generation

    Graph Learning for Task Planning (NeurIPS'2024)

    In language agents, available tasks naturally form a task graph, where nodes represent tasks and edges denote dependencies. Under such context, task planning involves selecting a path within this graph to fulfill user requests. We find that the bottleneck in LLMs' planning abilities lies in their limited understanding of the task graph. Therefore, we introduce GNNs as a simple fix, available in both training-free and training-required variants. Extensive experiments demonstrate that GNN-based methods surpass existing solutions even without training.


Selected Publications

Experience

  • generation

    Microsoft Research Asia
    Research Intern, Shanghai AI/ML Group Feb. 2024 - Jun. 2024

  • generation

    Microsoft
    Software Engineer Intern, Outlook Mobile Team Jul. 2020 - Sep. 2020


Selected Awards

  • Hong Kong PhD Fellowship (HKPFS), Hong Kong SAR 2024
  • National Scholarship for Graduate Student, Ministry of Education, China 2022 & 2023
  • ACM Web Conference Student Travel Award 2023
  • Second Class Scholarship for Outstanding Student, Fudan University 2018 & 2021
  • Second Prize of Undergraduate Mathematical Contest in Modeling, Shanghai, China (CUMCM) 2019
  • First Prize in National Olympiad in Mathematics in Provinces, Jiangsu, China 2016

Professional Services

  • Conference Reviewer: NeurIPS'2024 (Top Reviewer Award🏆) - 2025, ICLR'2025, ICML'2025, WWW'2024 Graph Foundation Model (GFM) Workshop, SIGKDD'2024 - 2025
  • Journal Reviewer: IEEE Transactions on Knowledge and Data Engineering (TKDE), Transactions on Machine Learning Research (TMLR)
  • Oral Presentation: WWW'2023

Miscellaneous

  • I love sports like swimming 🏊 and running 🏃. I also enjoy cooking Chinese food 🤣
  • During my undergraduate studies, I was interested in mobile app development (You can find all the source codes on my GitHub):

    Lose Weight, a Fluter App

    Hulv, a Mini-Program

    Chatroom, a Desktop App
  • I enjoy exploring the unknown and strive to keep moving forward on this path of discovery and learning ✨

Last updated at May 1, 2025 by Xixi