LLM agents for automated playtesting
Designing agent workflows that can explore game environments and produce useful behavioral evidence.
Research
Research
LLM agents, player modeling, and human-grounded evaluation for interactive entertainment systems.
Research Agenda
Xunzhuo's current research sits at the intersection of HCI, game AI, player modeling, and LLM agents. He is especially interested in how agent behavior can be evaluated against human traces, how automated playtesting can move beyond outcome-only metrics, and how future interactive entertainment systems can become more adaptive, explainable, and human-centered.
Player modeling and behavioral traces
Turning gameplay traces into interpretable patterns for player type reasoning and design feedback.
Human-grounded evaluation of AI agents
Comparing agent behavior against human traces instead of relying only on outcome metrics.
Current Research Projects
Public-safe summaries of active and planned research directions.
In progress 2025 - present · Researcher and system builder
LLM-Based Agents for Player-Type-Aware Automated Playtesting
Thesis direction
A research direction exploring how LLM-based agents can support automated playtesting while reflecting different player behavior patterns.
LLM Agents/Automated Playtesting/Player Modeling/Game AI
Working 2025 - present · Researcher
Human-Anchored Behavioral Comparison in Game Environments
Method project
A method-oriented project for comparing AI agent behavior with human traces across interactive game environments.
Behavior Traces/Human-Agent Comparison/Evaluation/HCI
Planned Ongoing · Researcher and builder
Trustworthy and Explainable LLM Agents for Interactive Systems
A broader research agenda around making LLM agents more interpretable, controllable, and useful in interactive entertainment and HCI contexts.
Explainability/Trust/Interactive Systems/Human-AI Interaction
Methods and Technical Strengths
Future Direction
Trustworthy and explainable agents for interactive systems
Making agent behavior legible enough that designers and researchers can trust what automated playtesting reports, and see why an agent acted the way it did.
Human-centered evaluation methods for agent behavior
Developing evaluation practices that compare agents against real human traces, rather than relying on outcome metrics alone.
Adaptive entertainment systems grounded in player evidence
Exploring how player modeling can drive interactive systems that adapt responsibly to how people actually play.