cs.LGNov 4, 2026

Where It Breaks: Diagnosing Perception, Memory, and Decision Failures of a Vision-Language Agent in Pac-Man

Authors: Xinglu Wang, Edward Wang, Yong Zhang, Zhenan Fan

Abstract

Training an agent to play Pac-Man with only up, down, left and right actions raises a basic question: when performance is poor, what should be improved? Difficulty can arise from reading the screen or from choosing actions over a long sequence. We examine these problems in a Qwen3.5-9B-based agent through separate perception checks and primitive-policy reinforcement-learning studies. In the policy study, which uses textual boards to remove image transcription from the decision interface, completion on eight held-out mazes increased from 0% before this optimization to 80.2% at a selected checkpoint, with three extra lives. The perception checks show that accurate static screenshot reading can coexist with errors during interaction. A separate policy comparison also shows that further optimization can accompany worse action choices. Together, these findings motivate concrete checks for post-training: evaluate perception in use, distinguish training success from full-game validation, and examine decisions and later performance rather than relying on one selected score. The study provides empirical lessons for improving future agent training and evaluation.

Figures & tables

Appendix figures & tables25 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

CardsList
  1. MNIST-PRO: MNIST is Back as a Partially Observable World for AI Agents

    Aug 31, 2026Vernon Toh, Navonil Majumder, Zhengyuan Liu +2Active PerceptionAI Agent Evaluation

  2. When History Fails to Become Experience: Action Calibration in Language Agents

    Oct 2, 2026Jingyu Liu, Zhiwen Wang, Yuxin Jing +2Memory-Augmented Language Model AgentsIn-Context Learning