About me

Hi, I’m Binyan! I am a Ph.D. candidate in Information Engineering at The Chinese University of Hong Kong, advised by Prof. Kehuan Zhang. Before that, I received my B.Sc. in Automation from the Qian Xuesen Honors College at Xi’an Jiaotong University and studied EECS as an exchange student at UC Berkeley.

My research asks how capable models become reliable, long-running agents. I make an agent’s operating state explicit, turn interaction traces into reusable experience and skills, and externally verify model behavior when internal signals cannot be trusted. I have led four accepted and four under-review CCF-A papers as first or co-first author.

Research interests

My research centers on long-horizon autonomous intelligence and its safety boundaries, including:

• Self-managed context and agentic RL: exposing token budget, recency, and working-state signals so agents can manage long trajectories and tool interactions.
• Experience and skill learning: separating temporary retrieval from durable memory, and distilling multi-agent workflows into efficient single-agent skills.
• Model-agnostic verification: using independent VLM/CLIP-based auditors to verify behavior without relying on a target model’s internal state.
• Autonomy risk and foundation-model security: studying transferable attacks, training-data integrity, generative backdoors, and the deployment economics of autonomous systems.

Selected papers

See the publications page for the full list. * indicates equal contribution.

Topic 1: Long-horizon agents, memory & skills

Under review

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

Binyan Xu, Haitao Li, Kehuan Zhang

Under review, 2026

arXivProjectCode

Under review

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

Binyan Xu, Dong Fang, Haitao Li, Kehuan Zhang

Under review, 2026

arXivCode

Under review

Contextual Agentic Memory is a Memo, Not True Memory

Binyan Xu*, Xilin Dai*, Kehuan Zhang

Under review, 2026

arXiv

Topic 2: External auditing & data integrity

ICML ’26

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

arXiv

ACM MM ’25

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

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

ACM Multimedia 2025, Oral Presentation

arXivACM DLCode

Topic 3: Open-model behavior & risk boundaries

CCS ’25

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, Oral Presentation

arXivACM DLCode

AAAI ’26

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, Oral Presentation

arXivCode

Projects

• VISTA: a self-managed context layer that gives agents proprioception over working memory. 22.7 → 50.7%
• AdaSkill: distilling multi-agent workflows into reusable single-agent skills. 1.4–15× cheaper
• Open-source research artifacts: reproducible code for agent memory, VLM auditing, backdoor defense, and transferable attacks.

Education

Experience

Honors & service

• Finalist, Mathematical Contest in Modeling (Top 2%).
• XJTU Honors Youth Program (approximately 100 admitted from 4,000+ applicants); Qian Xuesen Honors and Outstanding Graduate.
• First-class funding for the UC Berkeley exchange program.
• Microsoft Research Asia Stars of Tomorrow — Award of Excellence.
• Reviewer for ICML, ICLR, ACM MM, and AAAI; Lead TA supervising 40+ undergraduate capstone students.