Wendy Z.

Key Skills

Software

No software listed

Top Subject Matter

Multi-agent reinforcement learning

Adversarial evaluation for multi-agent systems

LLM evaluation

Top Data Types

Text

Top Task Types

RLHF

Question Answering

Red Teaming

Evaluation/Rating

Freelancer Overview

I am a Software Engineering undergraduate at South China University of Technology and will pursue a master’s degree in the United States. I have a strong background in programming, machine learning, and data analysis. Through my research on LLM-based multi-agent systems, I gained experience in AI training data and model evaluation, including reviewing outputs for safety, correctness, helpfulness, instruction-following, and robustness against prompt injection. I also categorized failure cases such as unsafe responses, malicious attacks, model replacement, communication delays, and task breakdowns. These experiences strengthened my attention to detail, ability to follow structured guidelines, and skill in documenting error patterns. My background in Python, machine learning, and JSON-style data helps me handle annotation tasks requiring both careful judgment and technical understanding.

Labeling Experience

Research position with Prof. Han Liu, MAGICS Lab @ Northwestern University (Evo-MARL/AdvEvo-MARL safety co-evolution)

TextRed Teaming

Conducted research to improve internalized safety in LLM-based multi-agent systems by co-training attackers and defenders in a multi-agent reinforcement learning setup. Implemented co-evolutionary training components that shape model behavior for safety, task performance, and response formatting. Evaluated adversarial scenarios to quantify attack success rates and measure downstream task utility. • Co-developed Evo-MARL and AdvEvo-MARL training frameworks for internalized safety. • Designed multi-objective reward shaping across safety, task utility, and response formatting. • Built baseline mechanisms for variance reduction via shared mean-return baselines among agents. • Reported evaluation results showing reduced attack success and improved benchmark performance.

2025 - 2025

Research position with Prof. Nuno Vasconcelos, SVCL Lab @ UC San Diego (robustness testing tool)

TextRed Teaming

Developed a robustness testing tool to automatically evaluate adversarial failure modes in distributed multi-agent systems. Focused on generating and assessing adversarial conditions across varied network topologies and communication patterns. Analyzed and quantified distinct vulnerability-driven breakdown modes in multi-agent task completion. • Built automated adversarial evaluation for distributed multi-agent robustness testing. • Identified failure modes such as backbone model replacement, malicious attacks, communication delay, and agent disconnection. • Demonstrated how these vulnerabilities can trigger complete task breakdown under adversarial conditions. • Supported robustness assessment through controlled experimental setups and measurements.

2024 - 2025

Education

Work History