MrNicholas Qi

My research centers on advancing trustworthy and effective AI systems, particularly in educational and multi-agent settings. My key research areas include:

1. Reinforcement Learning & Reward Hacking Mitigation

A major focus of my research is addressing reward hacking—a critical challenge in reinforcement learning where agents exploit poorly specified objectives. My work explores:

• Multi-objective reinforcement learning frameworks to align agent behavior with intended goals
• Detection and mitigation strategies for reward manipulation in complex environments
• Evaluation methodologies combining statistical testing (e.g., ANOVA, T-tests) with qualitative metrics such as reasoning quality and feedback utilization

I have proposed frameworks (e.g., RHRCLAW) that aim to systematically identify and reduce unintended optimization behaviors in both single-agent and multi-agent systems.

2. Multi-Agent Collaboration & LLM-Based Simulation

Nicholas develops architectures for collaborative AI systems where multiple agents interact to solve complex tasks. A representative system includes:

• Tutor Agent:
  Uses prompt-based reasoning to assess student learning progress, perform initial diagnosis, and generate structured feedback.
• Learner Agent:
  Leverages large language models to interpret tutoring guidelines and student learning records. It simulates student responses and engages in iterative dialogue with the tutor agent.

This system enables:

• Simulation of tutor–student interactions at scale
• Evaluation of diagnostic accuracy and feedback quality
• Iterative improvement of learning outcomes through multi-turn reasoning

3. AI in Education & Personalized Learning Systems

My broader research explores how LLMs and agent-based systems can:

• Provide personalized, adaptive feedback to learners
• Model student cognition and learning trajectories
• Enhance engagement through conversational and reflective learning modes

I have worked with datasets such as GSM8K and programming evaluation corpora to benchmark learning effectiveness across AI-generated and human-generated responses.