The research area of multimodal large language models (MLLMs) is currently advancing towards more robust and reliable systems, particularly in addressing the issue of hallucinations. Recent developments emphasize a shift from symptom-focused corrections to deeper understanding and mitigation of hallucination causes. Innovations include novel frameworks that integrate perception-level information with cognition-level commonsense knowledge, effectively bridging the gap between visual and textual inputs. Additionally, methods that manipulate model weights to suppress hallucinated features and enhance vision-aware attention heads are gaining traction. These approaches not only improve model accuracy but also maintain efficiency, often without additional inference costs. Token-level optimization strategies, leveraging visual-anchored rewards, are also emerging as powerful tools for aligning model outputs more closely with human preferences. Overall, the field is progressing towards more holistic and contextually accurate multimodal reasoning systems.