👋 I am currently pursuing an M.S. in Computer Science at Northeastern University. My research interests include NLP, information retrieval, Retrieval-Augmented Generation, recommender systems, and multi-agent systems. You can find my publications on my Google Scholar profile 
.
🎓 I received my B.S. in Computer Science from Harbin Engineering University. My research is carried out at the NEU-ModelBest Data Intelligence Joint Lab, jointly established by Northeastern University and ModelBest, under the guidance of Associate Professor Zhenghao Liu. I also collaborate closely with Qiyuan Lab 
, ModelBest 
, and TsinghuaNLP 
, and am mentored by Associate Researcher Yukun Yan.
🔥 News
- 2026.07: 🎉 We open-sourced StaffDeck
, an enterprise digital employee platform! - 2026.05: 🎉 We released PilotDeck
, a workspace-centric open-source agent operating system! - 2026.04: 🎉 Our paper MetaMem has been accepted at ACL 2026 (CCF-A)!
- 2026.02: 🎉 Our paper Knowledge Intensive Agents has been accepted by AI Open (IF=14.8)!
- 2026.01: 🎉 UltraRAG v3
reached #2 on GitHub Trending! - 2026.01: 🎉 We released UltraRAG v3
, making every step of the reasoning workflow visible!
- 2025.11: 🎉 Our paper LISRec has been accepted at KDD 2026 (CCF-A)!
- 2025.08: 🎉 Our paper LLMsPark has been accepted at EMNLP 2025 (CCF-B)!
- 2024.04: 🎉 We released heu-icicles
, an HEU course guide and resource-sharing platform. - 2024.02: 🎉 Our paper MMAD has been accepted at COLING 2024 (CCF-B)!
🔨 Projects
StaffDeck: Enterprise Digital Employee Platform
ModelBest, NEU-ModelBest Data Intelligence Joint Lab, THUNLP, OpenBMB, and AI9Stars
StaffDeck is an enterprise platform for building and managing digital employees. It turns professional experience, business processes, and decision criteria into digital employees that can work continuously, take over repetitive tasks, and preserve individual expertise as reusable, evolvable, and traceable organizational assets.
PilotDeck: A Workspace-Centric Open-Source Agent Operating System
THUNLP, ModelBest, OpenBMB, and AI9Stars
PilotDeck is an open-source agent operating system built around workspace-level isolation, white-box memory, smart routing, and always-on background execution for long-running productivity tasks.
UltraRAG 3.0: No More “Blind-Box” Development — Every Line of Reasoning Made Explicit
THUNLP, NEU-ModelBest Data Intelligence Joint Lab, OpenBMB, and AI9Stars
UltraRAG is a low-code MCP-based RAG framework that standardizes retrieval and generation components as MCP servers, with visual orchestration, evaluation, and demo deployment workflows.
📝 Publications
* indicates equal contribution, and † indicates corresponding author.
Test-Time Memory Construction
Zijun Wang*, Yifan Ji*, Haidong Xin, Zhenghao Liu†, Yukun Yan, Zulong Chen, Yu Gu, Ge Yu, Maosong Sun
T2Mem defers memory construction to inference, progressively retrieving episode metadata and raw interactions to build query-specific memory. It improves long-term dialogue question answering on LoCoMo and LongMemEval while reducing offline construction overhead.
MetaMem: Evolving Meta-Memory for Knowledge Utilization through Self-Reflective Symbolic Optimization
Haidong Xin*, Xinze Li*, Zhenghao Liu†, Yukun Yan†, Shuo Wang, Cheng Yang, Yu Gu, Ge Yu, Maosong Sun
MetaMem is a meta-memory framework that keeps its memory evolving through self-reflective symbolic optimization, enabling more effective knowledge use across long-horizon human-LLM interactions.
LISRec: Modeling User Preferences with Learned Item Shortcuts for Sequential Recommendation
Haidong Xin, Zhenghao Liu†, Sen Mei, Yukun Yan, Shi Yu, Shuo Wang, Zulong Chen, Yu Gu, Ge Yu, Chenyan Xiong
LISRec strengthens sequential recommendation by learning personalized semantic shortcuts from user-item interaction histories, capturing stable user preferences more reliably.
Knowledge Intensive Agents
Zhenghao Liu, Pengcheng Huang, Zhipeng Xu, Xinze Li, Shuliang Liu, Chunyi Peng, Haidong Xin, Yukun Yan†, Shuo Wang, Xu Han, Zhiyuan Liu†, Maosong Sun†, Yu Gu, Ge Yu
A comprehensive survey of Retrieval-Augmented Generation from an agentic perspective: it organizes knowledge-intensive agents into knowledge-acquisition and knowledge-utilization roles, and outlines future directions for jointly optimizing multi-agent RAG systems.
Adapting Language Models to Text Matching based Recommendation Systems
Haidong Xin, Sen Mei, Zhenghao Liu†, Xiaohua Li, Minghe Yu, Yu Gu, Ge Yu
By bringing language-model pretraining into recommendation, TASTE+ substantially improves the performance of text-matching-based recommendation.
LLMsPark: A Benchmark for Evaluating Large Language Models in Strategic Gaming Contexts
Junhao Chen, Jingbo Sun, Xiang Li, Haidong Xin, Yuhao Xue, Yibin Xu, Hao Zhao†
LLMsPark is a game-theoretic benchmark for evaluating the strategic behavior and decision-making of large language models in multi-agent environments, offering a new lens on model intelligence.
MMAD: Multi-modal Movie Audio Description
Xiaojun Ye, Junhao Chen, Xiang Li, Haidong Xin, Chao Li, Sheng Zhou†, Jiajun Bu
MMAD advances multi-modal movie audio description, bringing a richer movie-watching experience to people who are blind or have low vision.
Memory Shot for Long-Term Dialogue
Chunyi Peng*, Haidong Xin*, Xuanshuo Sheng, Xin Dai, Zhenghao Liu†, Shuo Wang, Yukun Yan, Zulong Chen, Yu Gu, Ge Yu
MemShot renders dialogue spans into structured visual memory units and uses the model’s visual reasoning to link key episodes, making long-term dialogue memory construction up to 70x faster.
CAGMamba: Context-Aware Gated Cross-Modal Mamba Network for Multimodal Sentiment Analysis
Minghai Jiao, Jing Xiao†, Peng Xiao, Ende Zhang, Shuang Kan, Wenyan Jiang, Jinyao Li, Yixian Liu, Haidong Xin
CAGMamba is a context-aware gated cross-modal Mamba network for multimodal sentiment analysis that models dialogue-level sentiment evolution and adaptively balances cross-modal fusion against modality-specific cues.
SEEK: Steering LLM Reasoning for RAG via Internal Reasoning Sketches
Xinze Li, Yuqing Lan, Zhenghao Liu†, Haidong Xin, Yukun Yan†, Shuo Wang, Zheni Zeng, Sen Mei, Ge Yu, Maosong Sun
SEEK steers LLM reasoning for RAG with internal reasoning sketches that guide iterative retrieval, organize knowledge along multiple dimensions, and lead to more accurate answers.
Puzzle Game: Prediction and Classification of Wordle Solution Words
Haidong Xin*†, Fang Wu*, Zhitong Zhou*
A detailed quantitative study of the Wordle game that uncovers the statistical patterns behind its solution words.
🏆 Awards
- 2025.11 🥇 First-Class Excellent Student Scholarship.
- 2024.11 🥇 First-Class Excellent Student Scholarship.
- 2023.07 🥈 National Second Prize, Chinese Collegiate Computing Competition (4C 2023).
- 2022.11 🥇 First Prize, China Undergraduate Mathematical Contest in Modeling (CUMCM 2022).
- 2022.10 🥈 Second Prize, Mathematical Modeling Competition of the Three Northeastern Provinces.
- 2022.05 🥇 First-Class Excellent Student Scholarship.
- 2021.05 🥇 First-Class Excellent Student Scholarship.
📖 Education
- 2024.09 - present, M.S. in Computer Science, School of Computer Science and Engineering, Northeastern University
- 2020.09 - 2024.06, B.S. in Computer Science, College of Computer Science and Technology, Harbin Engineering University
- 2017.09 - 2020.06, Yichun No. 1 Middle School
💬 Media Mentions
- 2026.10, 【TsinghuaNLP】成果 | PilotDeck 2.0 开源：一键上手，长难任务高效稳定交付
- 2026.07, 【量子位】StaffDeck 正式开源：面向企业的数字员工构建与管理平台
- 2026.05, 【TsinghuaNLP】成果 | PilotDeck：一人，一桌面，有序管理一支 AI 智能体团队
- 2026.04, 【TsinghuaNLP】成果 | MetaMem：赋予大模型自我进化的“元记忆”
- 2026.01, 【TsinghuaNLP】成果 | UltraRAG 3.0：拒绝“盲盒”开发，让每一行推理逻辑都清晰可见
- 2025.11, 【TsinghuaNLP】成果 | UltraRAG 2.1：强化知识接入与多模态支持，完善统一评估体系
- 2025.08, 【TsinghuaNLP】成果 | UltraRAG 2.0：几十行代码实现高性能 RAG，让科研专注思想创新
💻 Research and Internships
- 2024.06 - present, Beijing ModelBest Intelligent Technology Co., Ltd., Beijing.
- 2024.04 - 2026.10, Qiyuan Lab, Beijing.
- 2024.10 - 2025.02, Language Technologies Institute (LTI), Carnegie Mellon University, Remote.
- 2023.10 - present, NEU-ModelBest Data Intelligence Joint Lab, Northeastern University, Shenyang.