Nihon University

Synergizing Code Coverage and Gameplay Intent: Coverage-Aware Game Playtesting with LLM-Guided Reinforcement Learning

The SMART (Structural Mapping for Augmented Reinforcement Testing) framework introduces a novel methodology for automating game quality assurance by bridging the gap between low-level structural code analysis and high-level functional gameplay intent.

• The SMART framework achieves a 94% branch coverage rate and a 98% task completion rate by integrating LLM-guided reinforcement learning with structural code analysis.
• The methodology bridges the gap between low-level code changes and high-level gameplay intent by using LLMs to decompose updates into actionable subgoals and structural anchors.
• Experimental evaluations in Overcooked and Minecraft demonstrate that combining semantic subgoals with structural rewards is essential, as neither approach is sufficient in isolation.