VideoLLMs excel at video understanding but still suffer from action hallucination. Prior work mainly analyzes it at the frame level, overlooking patch-level causes. We find that action evidence concentrates in dynamic patches, whose answer-related signals are stronger in non-hallucinated cases. We thus propose DEGD, training-free decoding that emphasizes dynamic evidence, and ACE, which evaluates action hallucination in generative responses.

Sunhee Jo received a B.S. degree in Electronics Engineering in 2025. She is currently pursuing an M.S. degree in AI Imaging at the Graduate School of Advanced Imaging Science, Multimedia and Film, Chung-Ang University. Her current research interests include multimodal learning and Video Question Answering (VideoQA).