Wujiang Xu' s homepage

Wujiang Xu 徐武将

WechatIMG794.jpg

I am a third-year Ph.D. student at Rutgers University working with Prof. Dimitris N. Metaxas. Previously, I worked with Prof. Yongfeng Zhang (May 2024–June 2025). Before joining Rutgers, I was a Senior Machine Learning Engineer at Ant Group, with MLE internship experience at Tencent, ByteDance, and Ant Group. I am currently a Research Scientist Intern at Meta Superintelligence Labs (MSL) (May–Dec 2026), after a previous research internship at Meta (May–Aug 2025).

My research focuses on LLM agents that remember, adapt, and improve over time: agentic memory, self-evolving agents and agentic reinforcement learning, agent harness and infrastructure, and personalized agents. I also study how foundation models work internally. If you are interested in my research or potential collaboration opportunities, please feel free to contact me by email.

I am currently looking for full-time positions and research internships — please reach out if there is a fit!

news

Aug 01, 2026 Our paper AEL: Evolving Agent Harness in Open-Ended Environments has been accepted by Findings of EMNLP 2026. Congratulations to all collaborators!
May 01, 2026 Joined Meta Superintelligence Labs (MSL) as a Research Scientist Intern (May–Dec 2026), after a previous research internship at Meta (May–Aug 2025).
Oct 17, 2025 Honored as a Top Reviewer for NeurIPS 2025.
Oct 11, 2025 Received NeurIPS 2025 Scholar Award.
Sep 18, 2025 A-mem paper has been accepted by NeurIPS 2025.

research

How should LLM agents store, organize, and evolve what they experience? I build self-organizing and structured memory systems (A-Mem, GAM, ChronoMem) and long-horizon, multimodal environments for evaluating them (MemGym, MemEye).

  1. amem.png
    A-mem: Agentic Memory for LLM Agents
    Wujiang Xu, Zujie Liang , Kai Mei , and 3 more authors
    In Advances in Neural Information Processing Systems (NeurIPS 2025) , 2025
  2. memgym.png
    MemGym: A Long-Horizon Memory Environment for LLM Agents
    Wujiang Xu, Yu Wang , Kai Mei , and 8 more authors
    arXiv preprint arXiv:2605.20833, 2026
  3. ACL
    GAM: Hierarchical Graph-based Agentic Memory for LLM Agents
    Zhaofen Wu , Hanrong Zhang , Fulin Lin , and 9 more authors
    In Annual Meeting of the Association for Computational Linguistics (ACL 2026) , 2026
  4. arXiv
    MemEye: A Visual-Centric Evaluation Framework for Multimodal Agent Memory
    Minghao Guo , Qingyue Jiao , Zeru Shi , and 14 more authors
    arXiv preprint arXiv:2605.15128, 2026
  5. arXiv
    ChronoMem: Version Control and Semantic Rollback for Large Language Model Agent Memory
    Yongye Su , Wujiang Xu, Chaoji Zuo , and 1 more author
    arXiv preprint arXiv:2607.27773, 2026
  6. TMLR
    Rethinking Memory Mechanisms of Foundation Agents in the Second Half: A Survey
    Wei-Chieh Huang , Weizhi Zhang , Yueqing Liang , and 57 more authors
    Transactions on Machine Learning Research (TMLR), 2026

Agents that keep improving from their own experience: stable multi-turn reinforcement learning for LLM agents (EPO), agents that evolve their own harness in open-ended environments (AEL), and introspective search for agentic problem solving (I-MCTS).

  1. ael.png
    AEL: Evolving Agent Harness in Open-Ended Environments
    Wujiang Xu, Jiaojiao Han , Minghao Guo , and 4 more authors
    In Findings of the Association for Computational Linguistics: EMNLP 2026 , 2026
  2. epo.png
    EPO: Entropy-regularized Policy Optimization for LLM Agents Reinforcement Learning
    Wujiang Xu, Wentian Zhao , Zhenting Wang , and 6 more authors
    arXiv preprint arXiv:2509.22576, 2025
  3. EACL Findings
    I-MCTS: Enhancing Agentic AutoML via Introspective Monte Carlo Tree Search
    Zujie Liang , Feng Wei , Wujiang Xu, and 3 more authors
    In Findings of the European Chapter of the Association for Computational Linguistics (EACL 2026) , 2026

The system layer around the model that turns an LLM into a working agent: agent operating systems (AIOS), semantic file systems (LSFS), budget-aware multi-LLM routing (OmniRouter), and automatic prompt engineering (APEER).

  1. COLM
    AIOS: LLM Agent Operating System
    Kai Mei , Xi Zhu , Wujiang Xu, and 7 more authors
    In Conference on Language Modeling (COLM 2025) , 2025
  2. ICLR
    From Commands to Prompts: LLM-based Semantic File System for AIOS
    Zeru Shi , Kai Mei , Mingyu Jin , and 9 more authors
    In International Conference on Learning Representations (ICLR 2025) , 2025
  3. SIGKDD Explor.
    OmniRouter: Budget and Performance Controllable Multi-LLM Routing
    Kai Mei , Wujiang Xu, Minghao Guo , and 2 more authors
    ACM SIGKDD Explorations Newsletter, 2025
  4. WWW Companion
    APEER: Automatic Prompt Engineering Enhances Large Language Model Reranking
    Can Jin , Hongwu Peng , Shiyu Zhao , and 7 more authors
    In Companion Proceedings of the ACM Web Conference 2025 , 2025

Agents that understand and act on behalf of individual users: user-side agents that shield users from recommender systems (iAgent), long-behavior user modeling (PersonaX), simulating individuals from longitudinal personal data, and recommending agents themselves (AgentSelect).

  1. ACL Findings
    iAgent: LLM Agent as a Shield between User and Recommender Systems
    Wujiang Xu, Yunxiao Shi , Zujie Liang , and 6 more authors
    In Findings of the Association for Computational Linguistics ACL 2025 , 2025
  2. ICML
    AgentSelect: Benchmark for Narrative Query-to-Agent Recommendation
    Yunxiao Shi , Wujiang Xu, Tingwei Chen , and 7 more authors
    In International Conference on Machine Learning (ICML 2026) , 2026
  3. ACL Findings
    PersonaX: A Recommendation Agent Oriented User Modeling Framework for Long Behavior Sequence
    Yunxiao Shi , Wujiang Xu, Zeqi Zhang , and 3 more authors
    In Findings of the Association for Computational Linguistics ACL 2025 , 2025
  4. SIGIR
    Individual Turing Test: A Case Study of LLM-based Simulation Using Longitudinal Personal Data
    Minghao Guo , Ziyi Ye , Wujiang Xu, and 3 more authors
    In International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR 2026), Short Paper , 2026
  5. EMNLP
    Answering Narrative-Driven Recommendation Queries via a Retrieve–Rank Paradigm and the OCG-Agent
    Yunxiao Shi , Haoning Shang , Xing Zi , and 3 more authors
    In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP 2025) , 2025
  6. slmrec.png
    SLMRec: Distilling Large Language Models into Small for Sequential Recommendation
    Wujiang Xu, Qitian Wu , Zujie Liang , and 5 more authors
    In International Conference on Learning Representations (ICLR 2025) , 2025
  7. FnT
    Recommender Systems Meet Large Language Model Agents: A Survey
    Xi Zhu , Yu Wang , Hang Gao , and 9 more authors
    Foundations and Trends\textregistered in Privacy and Security, 2025

Looking inside LLMs to understand how they work and where they fail: massive values in attention, out-of-distribution representations, agentic interpretation of SAE features, and moral evaluation, as well as graph and multimodal foundation models.

  1. ICML
    Massive Values in Self-Attention Modules are the Key to Contextual Knowledge Understanding
    Mingyu Jin , Kai Mei , Wujiang Xu, and 5 more authors
    In International Conference on Machine Learning (ICML 2025) , 2025
  2. NeurIPS
    Farther the Shift, Sparser the Representation: Analyzing OOD Mechanisms in LLMs
    Mingyu Jin , Yutong Yin , Jingcheng Niu , and 7 more authors
    In Advances in Neural Information Processing Systems (NeurIPS 2026) , 2026
  3. EACL Industry
    SAGE: An Agentic Explainer Framework for Interpreting SAE Features in Language Models
    Jiaojiao Han , Wujiang Xu, Mingyu Jin , and 1 more author
    In Conference of the European Chapter of the Association for Computational Linguistics (EACL 2026), Industry Track , 2026
  4. SIGKDD Explor.
    MoralBench: Moral Evaluation of LLMs
    Jianchao Ji , Yutong Chen , Mingyu Jin , and 3 more authors
    ACM SIGKDD Explorations Newsletter, 2025
  5. EMNLP
    LLM as GNN: Graph Vocabulary Learning for Text-Attributed Graph Foundation Models
    Xi Zhu , Haochen Xue , Ziwei Zhao , and 7 more authors
    In Conference on Empirical Methods in Natural Language Processing (EMNLP 2026) , 2026
  6. ICML
    Graph4MM: Weaving Multimodal Learning with Structural Information
    Xuying Ning , Dongqi Fu , Tianxin Wei , and 2 more authors
    In International Conference on Machine Learning (ICML 2025) , 2025

Earlier work on cross-domain recommendation and computer vision can be found in the full publication list.