PhD Thesis Proposal: Qintong Xie

ZOOM LINK

"Structured Reasoning for Trustworthy Autonomous Decision-Making"

Abstract

Modern autonomous systems increasingly operate in environments where decisions are sequential, uncertain, and exposed to strategic or adversarial behavior. In cybersecurity, defenders must adapt to attackers under partial observability. In software engineering, maintainers increasingly rely on LLM-assisted code review while facing authors who may submit low-quality, misleading, or adversarial pull requests. In robotics, agents must reason about physical dynamics, long-horizon task structure, and safety constraints before acting. Although machine learning models have achieved strong empirical performance across these domains, purely end-to-end approaches often struggle with robustness, interpretability, safety assurance, and generalization under distribution shift.

This thesis proposes that trustworthy autonomous decision-making requires structured reasoning: explicit game-theoretic, symbolic, and model-based structure integrated with neural learning. The first part develops equilibrium-informed multi-agent reinforcement learning methods for adversarial and competitive environments. Building on Nash Q-Networks for cybersecurity and Deep Nash Q-Networks for partially observable n-player games, this work studies how agents can learn policies guided by equilibrium computation while managing the scalability limits of exact game solving.

The second part introduces a new governance formulation for LLM-assisted GitHub pull request review. Rather than assuming cooperative authors, this work treats authors as uncontrollable and potentially adversarial. The controllable object is the reviewer--maintainer pipeline: a governance mechanism that decides when to merge, reject, request revision, or escalate to human review. The goal is to design mechanisms that remain safe under worst-case author behavior by reducing harmful merges while preserving useful automation.

The third part studies structured reasoning in embodied decision-making. Building on safe multi-robot control, latent world models, and neuro-symbolic robotics, this thesis investigates how robots can combine predictive dynamics, symbolic task structure, and safety constraints to solve long-horizon tasks more reliably and efficiently.

Across these directions, the central argument is that autonomous agents should reason over incentives, uncertainty, dynamics, constraints, and risk. The expected contribution is a unified framework for designing AI systems that are more robust, interpretable, and trustworthy in both digital and physical environments.

Thesis Committee

• Peter Chin (Chair)
• Wesley Marrero
• Bryce Ferguson

Contact

For more information, contact Thayer Registrar at thayer.registrar@dartmouth.edu .