Binyan Xu 徐彬琰

论文

4 篇 CCF-A 已录用、4 篇在投，均为第一作者或共同第一作者工作。* 表示共同贡献。

在投LLM Agents Are Latent Context Managers: Eliciting Self-Managed Context via State Proprioception 论文插图

LLM Agents Are Latent Context Managers: Eliciting Self-Managed Context via State Proprioception

Binyan Xu, Haitao Li, Kehuan Zhang

在投，2026

通过 typed / addressable context layer 与 context-tool 动作空间显式暴露工作状态。在 75 个长程任务上将 Gemini-3-Flash 成功率从 22.7% 提升至 50.7%，并验证其与 GRPO 后训练互补。

arXiv项目主页代码
在投From Multi-Agent to Single-Agent: When Is Skill Distillation Beneficial? 论文插图

From Multi-Agent to Single-Agent: When Is Skill Distillation Beneficial?

Binyan Xu, Dong Fang, Haitao Li, Kehuan Zhang

在投，2026

提出 Metric Freedom，判断何时可将协作与评测闭环内化为单智能体技能；在 4 项任务、11 个数据集上以 1.4–15× 更低成本达到或超过原系统。

arXiv代码
在投Contextual Agentic Memory is a Memo, Not True Memory 论文插图

Contextual Agentic Memory is a Memo, Not True Memory

Binyan Xu*, Xilin Dai*, Kehuan Zhang

在投，2026

通过可控消融与统计分析指出，许多检索式“记忆”更接近临时上下文扩展，并围绕可复用、可泛化经验重新界定智能体记忆的评测目标。

arXiv
ICML ’26From Internal Diagnosis to External Auditing: A VLM-Driven Paradigm for Data-Free Online Backdoor Defense 论文插图

From Internal Diagnosis to External Auditing: A VLM-Driven Paradigm for Data-Free Online Backdoor Defense

Binyan Xu, Xilin Dai, Fan Yang, Di Tang, Kehuan Zhang

ICML 2026，Poster

用独立视觉语言模型在线核验预测，将防御从不可信模型的内部诊断转向模型无关的外部语义审计。

arXiv
ACM MM ’25CLIP-Guided Backdoor Defense through Entropy-Based Poisoned Dataset Separation 论文插图

CLIP-Guided Backdoor Defense through Entropy-Based Poisoned Dataset Separation

Binyan Xu, Fan Yang, Xilin Dai, Di Tang, Kehuan Zhang

ACM Multimedia 2025，口头报告

利用 CLIP 语义信号与熵评分分离毒化 / 干净样本，并指导模型修复，在多种后门攻击下保持干净准确率。

arXivACM DL代码
CCS ’25One Surrogate to Fool Them All: Universal, Transferable, and Targeted Adversarial Attacks with CLIP 论文插图

One Surrogate to Fool Them All: Universal, Transferable, and Targeted Adversarial Attacks with CLIP

Binyan Xu, Xilin Dai, Di Tang, Kehuan Zhang

ACM CCS 2025，口头报告

以一个公开 CLIP 模型作为通用替身，在无需查询受害模型的情况下，构造跨模型、跨任务、多模态系统与真实黑盒服务的定向迁移攻击。

arXivACM DL代码
AAAI ’26Breaking the Stealth-Potency Trade-off in Clean-Image Backdoors with Generative Trigger Optimization 论文插图

Breaking the Stealth-Potency Trade-off in Clean-Image Backdoors with Generative Trigger Optimization

Binyan Xu, Xilin Dai, Fan Yang, Di Tang, Kehuan Zhang

AAAI 2026，口头报告

通过生成式触发器优化构造兼具隐蔽性与攻击效力的 clean-image 后门，并扩展至分类、回归和分割任务。

arXiv代码
在投When Agent Automation Becomes Profitable: Quantifying and Insuring Autonomous AI Risk through Trace-Economic Underwriting 论文插图

When Agent Automation Becomes Profitable: Quantifying and Insuring Autonomous AI Risk through Trace-Economic Underwriting

Binyan Xu, Xilin Dai, Fan Yang, Kehuan Zhang

在投，2026

将智能体工具轨迹映射为任务损失、控制信号与保险式定价，刻画在失败风险下自主执行何时具备经济可行性。

arXiv