Skip to content跳转到正文

About关于我

I'm Yijie Lu (陆一杰), a Ph.D. student at the School of Computer Science, Shanghai Jiao Tong University (SJTU), supervised by Prof. Zhuosheng Zhang. I received my B.E. in Cyberspace Security from Wuhan University (WHU) in 2026.

My research focuses on building agents that improve autonomously and operate reliably.

我是陆一杰(Yijie Lu),目前在上海交通大学计算机学院攻读博士学位,导师是张倬胜教授。我于 2026 年在武汉大学获得网络空间安全专业工学学士学位。

我的研究关注能够自主改进、可靠运行的智能体。

Research interests研究方向与兴趣

  • Self-Evolving Agents智能体自主演化Learning from experience to improve agent capabilities and behavior.从经验中学习,持续改进智能体的能力与行为。
  • Agent Reliability智能体可靠性Understanding failures and making agents dependable in real environments.理解失败机制,提升智能体在真实环境中的可靠性。
  • Agentic Post-Training智能体后训练Post-training for effective decision-making and interaction.通过后训练提升智能体的决策与交互能力。

I welcome discussions and collaborations. Feel free to get in touch.欢迎交流与合作,感兴趣的话可以通过邮件联系我。

You can also reach me on WeChat: SJTUcs0416.也欢迎添加微信交流:SJTUcs0416。

Seek truth through causes; renew through reflection; unite knowledge and action.循因求真,自省日新,知行合一。

News近况

  1. Became a core contributor and maintainer of REA, an open-source project for reverse engineering with AI agents.

    成为 AI 智能体逆向工程开源项目 REA 的核心贡献者和维护者。

  2. Invited to attend NVIDIA China Developer Day in Suzhou, Jiangsu. Looking forward to meeting you in Suzhou!

    受邀参加 NVIDIA 中国开发者日,将赴江苏苏州与开发者交流。期待在苏州相遇!

  3. Invited to an AI-focused doctoral exchange program in Zhejiang, with activities in Hangzhou. Looking forward to meeting you in Hangzhou!

    受邀参加 浙江省人工智能领域博士浙江行活动,将赴浙江杭州交流学习。期待在杭州相遇!

  4. Became a core contributor and maintainer of MLNLP-World / Paper-Writing-Tips, helping automate pre-submission paper checks.

    成为 MLNLP-World / Paper-Writing-Tips 的核心贡献者和维护者,支持论文投稿前的自动化检查。

  5. Our paper Communication Policy Evolution (CPE) has been accepted to EMNLP 2026. See you in Budapest!

    论文 Communication Policy Evolution(CPE) 中稿 EMNLP 2026,布达佩斯见!

  6. Our paper EVA has been accepted to ACL 2026. See you in San Diego!

    论文 EVA 中稿 ACL 2026,圣迭戈见!

Publications论文与研究

EMNLP 2026Accepted已接收

Communication Policy Evolution for Proactive LLM Agents

Xinbei Ma, Jiyang Qiu, Yao Yao, Zheng Wu, Yijie Lu, Xiangmou Qu, Jiaxin Yin, Xingyu Lou, Jun Wang, Weiwen Liu, Weinan Zhang, Zhuosheng Zhang, Hai Zhao

  • Self-evolution自主演化
  • Proactive agents主动式智能体
  • Communication沟通策略

We study how proactive agents communicate through text and structured interfaces, then evolve their communication policies through rollout and prompt refinement, without modifying model parameters.我们研究主动式智能体如何通过文本与结构化界面进行沟通,并通过任务执行与提示词优化演化沟通策略,无需修改模型参数。

Paper论文
Abstract摘要

LLM agents have rapidly evolved into autonomous systems, yet a persistent information gap remains between users and agents: communication is costly, while users' identical preferences further limit information exchange. To investigate how agents should communicate across modalities, this paper formalizes Communication Policy, establishes textual and UI-based policies, and then evaluates communication policies across diverse environments, personas, and model combinations. Building on information asymmetry for proactive agents, we set up two complementary settings, User--Agent and Planner--Executor. Experimental results reveal complementary strengths between interaction channels: text-based interaction often facilitates task performance, while structured UI improves agents' response quality and persona compliance. Motivated by that, a hybrid method that combines these advantages, we further propose Communication Policy Evolution (CPE), a self-evolution framework for refining communication policies through rollout and prompt-level evolution. Without model modification, CPE achieves the best task success across multiple settings using prompt refinement alone. Our findings identify communication behavior as a critical yet underexplored design dimension for LLM agents.

大语言模型智能体正逐步发展为自主系统,但用户与智能体之间仍存在信息鸿沟:沟通具有成本,用户偏好也会限制信息交换。为研究智能体如何通过不同模态进行沟通,本文形式化定义了沟通策略(Communication Policy),建立了基于文本和用户界面的策略,并在不同环境、用户角色及模型组合下进行评估。围绕主动式智能体中的信息不对称问题,我们构建了用户—智能体(User–Agent)与规划者—执行者(Planner–Executor)两种互补场景。实验发现,两类交互渠道各有所长:文本交互通常有利于任务完成,而结构化界面能够提升回复质量及对用户角色设定的遵循程度。在此基础上,我们提出沟通策略演化(Communication Policy Evolution,CPE),通过任务执行轨迹与提示词层面的演化持续改进沟通策略。无需修改模型参数,CPE 仅通过优化提示词便在多个场景下取得了最佳任务成功率。研究表明,沟通行为是大语言模型智能体设计中重要但尚未得到充分研究的维度。

BibTeX
@misc{ma2026communicationpolicyevolution,
  title = {Communication Policy Evolution for Proactive LLM Agents},
  author = {Xinbei Ma and Jiyang Qiu and Yao Yao and Zheng Wu and Yijie Lu and Xiangmou Qu and Jiaxin Yin and Xingyu Lou and Jun Wang and Weiwen Liu and Weinan Zhang and Zhuosheng Zhang and Hai Zhao},
  year = {2026},
  eprint = {2606.14314},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2606.14314}
}

ACL 2026Accepted已接收

EVA: Evolving Semantic Adversaries for Red-Teaming GUI Agents Against Environmental Injection Attacks

Yijie Lu, Manman Zhao, Tianjie Ju, Zihe Yan, Xinbei Ma, Yuan Guo, Daizong Ding, Gongshen Liu, Zhuosheng Zhang

  • Agent reliability智能体可靠性
  • GUI agentsGUI 智能体
  • Red-teaming红队测试

We identify semantic deception as a key vulnerability in GUI agents and introduce an evolutionary red-teaming framework that discovers reusable attack patterns, reaching up to 85% attack success across five victim agents.我们发现语义欺骗是 GUI 智能体的一项关键脆弱性,并提出演化式红队框架以挖掘可复用的攻击模式,在五种目标智能体上最高达到 85% 的攻击成功率。

Paper论文
Abstract摘要

Graphical User Interface (GUI) agents powered by Multimodal Large Language Models (MLLMs) are increasingly deployed yet vulnerable to Environmental Injection Attacks (EIAs).However, current red-teaming methods are hindered by prohibitive computational costs and limited adaptability.A fundamental question remains unaddressed: does the bottleneck of attack success lie in visual perception or semantic understanding? Through controlled experiments, we observe that semantic deception, rather than visual appearance, serves as the primary determinant of attack success. Based on this insight, we introduce EVA, an evolutionary framework that evolves adversarial payloads exclusively within the semantic dimension.EVA employs a discovery-deployment framework to mine linguistic vulnerability patterns and distill them into generalizable rules.Experimental results across five representative victim agents demonstrate that EVA achieves up to 85% attack success rate, evolving benign seeds into successful attacks within only 1.18 to 1.71 iterations.This rapid convergence uncovers a dense semantic attack space in the model’s latent representation, unveiling a critical alignment paradox: the instruction-following capabilities reinforced by alignment training render agents inherently susceptible to authoritative, semantically deceptive environmental cues.

基于多模态大语言模型的图形用户界面(GUI)智能体正在得到广泛应用,但容易受到环境注入攻击(Environmental Injection Attacks,EIAs)。现有红队测试方法面临计算成本高、适应性不足的问题。攻击成功的关键究竟在于视觉感知还是语义理解?通过受控实验,我们发现,相比视觉外观,语义欺骗是决定攻击成功的主要因素。基于这一发现,我们提出 EVA,一个仅在语义层面演化对抗载荷的框架。EVA 采用发现—部署流程,挖掘语言层面的脆弱性模式,并将其提炼为可泛化的规则。在五种代表性目标智能体上的实验中,EVA 的攻击成功率最高达到 85%,平均仅需 1.18 至 1.71 次迭代即可将良性初始样本演化为成功攻击。这种快速收敛揭示了模型潜在表示中密集的语义攻击空间,也揭示出一个关键的对齐悖论:对齐训练强化的指令遵循能力,使智能体容易受到具有权威性、语义欺骗性的环境线索影响。

BibTeX
@misc{lu2025evasemanticadversaries,
  title = {EVA: Evolving Semantic Adversaries for Red-Teaming GUI Agents Against Environmental Injection Attacks},
  author = {Yijie Lu and Manman Zhao and Tianjie Ju and Zihe Yan and Xinbei Ma and Yuan Guo and Daizong Ding and Gongshen Liu and Zhuosheng Zhang},
  year = {2025},
  eprint = {2505.14289},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2505.14289}
}

Internships实习

I am currently seeking internship opportunities in agent post-training and RIS, and I look forward to connecting and collaborating with more people in these areas.我目前正在寻找智能体后训练和 RIS 相关的实习机会,也期待与更多人交流与合作。

Education教育经历

  1. Sep 2026–present2026.09 至今

    Shanghai Jiao Tong University上海交通大学

    Ph.D. in Cyberspace Security网络空间安全博士在读

  2. Aug 2022–Jun 20262022.08–2026.06

    Wuhan University武汉大学

    B.E. in Cyberspace Security网络空间安全工学学士

Honors奖学金与荣誉

  • Outstanding Graduate of WHU武汉大学优秀毕业生Wuhan University武汉大学
  • Lei Jun Computer Science Undergraduate Scholarship雷军计算机本科生奖学金Wuhan University & Xiaomi Inc.武汉大学、小米集团
  • First Class Scholarship of WHU武汉大学一等奖学金Wuhan University武汉大学
  • Merit Student武汉大学三好学生Wuhan University武汉大学
  • LvMeng Scholarship绿盟奖学金Wuhan University武汉大学
  • Advanced Individual in Scientific and Technological Innovation科技创新先进个人Wuhan University武汉大学
  • Lei Jun Fund for Computer Science Innovation and Development雷军计算机创新与发展资助基金Funding recipient获资助

Service & Activities活动与服务

  • Jun–Jul 20252025.06–2025.07

    Teaching Assistant助教

    WHU-Jisuanke Joint Practical Training Course “Security Maker Practice”武汉大学—计蒜客联合实训课程「安全创客实践」

    Awarded “Top TA” for contributions to teaching, exercise explanation, and Q&A sessions.参与教学、习题讲解和答疑,获评「Top TA」。

  • Jul–Aug 20242024.07–2024.08

    Teaching Assistant助教

    WHU-Jisuanke Joint Practical Training Course “Security Maker Practice”武汉大学—计蒜客联合实训课程「安全创客实践」

    Awarded “Excellent TA” for contributions to the one-month practical training course.参与为期一个月的实训教学,获评「优秀助教」。

  • Aug 2022–Jul 20242022.08–2024.07

    Main Member骨干成员

    SITS Skating ClubSITS 轮滑社

  • Aug 2022–Jun 20242022.08–2024.06

    Member成员

    Ziqiang Student Network Culture Studio自强学生网络文化工作室