cv

Applied Scientist Intern

Langboat Technology

Remote applied scientist intern focusing on language model agents.

• Studied topics on Computer-Using Agents including long-horizonal compositional tasks and personalization under the supervision of Prof Ming Zhou (ACL Fellow)
• Involved in building agentic systems for enterprise clients
• Hosted multiple tech salons, prepared 2 patents (1 already issued) and wrote several tech blogs

Research Intern

AI Security Lab, SJTU

Research on autonomous GUI agent and agent safety under Prof Zhuosheng Zhang.

• Studied topics about autonomous GUI agent, agent safety etc

Undergraduate (Yongqiang Honor Class)

Shanghai Jiao Tong University

Computer Science and Technology

High School

Zhenhai High School, Zhejiang Province

Advanced Class of Sciences

National Scholarship

Ministry of Education of China

(top 0.2% nationwide)

Fan-Xuji Scholarship

Shanghai Jiao Tong University

(15 excellent second-year undergraduates in SJTU)

Yongqiang Encouragement Scholarship

Shanghai Jiao Tong University

(top 30% elite students in CS department)

SJTU Academic Scholarship

Shanghai Jiao Tong University

(top 10%)

Atomic-to-Compositional Generalization for Mobile Agents with A New Benchmark and Scheduling System

arXiv preprint

Introduced a comprehensive benchmark UI-NEXUS for mobile agents on compositional tasks and proposed a scheduling system for atomic-to-compositional generalization for existing mobile (computer-using) agents.

Universal Confidence Integration Framework for Adaptive Interaction in Computer-Using Agents

under review

A framework for blending agents for adaptive, human-in-the-loop human-agent-device co-interaction. PAMI submission.

EVA: Red-Teaming GUI Agents via Evolving Indirect Prompt Injection

arXiv preprint

We propose EVA, a red-teaming framework for indirect prompt injection which transforms the attack into a closed-loop optimization by continuously monitoring an agent's attention distribution over the GUI and updating adversarial cues—keywords, phrasing, and layout—in response.

BeHonest: Benchmarking Honesty of Large Language Models

arXiv preprint

In this paper, we introduce BeHonest, a pioneering benchmark specifically designed to assess honesty in LLMs comprehensively. BeHonest evaluates three essential aspects of honesty: awareness of knowledge boundaries, avoidance of deceit, and consistency of responses.

Plan, Generate and Match: Scientific Workflow Recommendation with Large Language Models

ICSOC 2023

We propose 'Plan, Generate and Match' (PGM), a scientific workflow recommendation method leveraging LLMs. PGM consists of three stages: utilizing LLMs for planning by generating a user query, generating a structured workflow specification from Solution Steps, and using these plans and specifications to match with candidate workflows.

Model construction method, program product and storage medium for GUI agent

Patent No: CN119576470B

Status: Issued

A personalization technology for MLLM-powered GUI agents.