Vision–Language Synergistic Intelligence in Multimodal Large Language Models (MLLMs): Capability Boundaries, Hallucination Mitigation, and Reasoning Enhancement
Keywords:
Multimodal Large Language Model, Vision-Language Collaboration, Capability Boundary, Illusion Suppression, Reasoning EnhancementAbstract
Multimodal large language models, by connecting visual representation and language generation, have expanded the task scope of image question answering, document understanding, and complex reasoning. However, their output is still constrained by visual information loss, language prior bias, and the accumulation of reasoning errors. Based on a selection of representative literature, this paper reviews the visual-language collaborative mechanism from three aspects: capability boundary, illusion suppression, and reasoning enhancement. It compares the effects and limitations of training alignment, decoding intervention, structured reasoning, and active visual search. The comprehensive analysis suggests that overall task performance, fidelity to image evidence, and reliability of the reasoning process should be evaluated separately; increasing model size or reasoning length is insufficient to guarantee simultaneous improvement in all three. Future research needs to strengthen fine-grained evidence localization, process supervision, dynamic computational allocation, and cross-scenario verification, and establish an evaluation system that balances accuracy, traceability, and cost.Downloads
Published
2026-09-22
How to Cite
Sun, Y. (2026). Vision–Language Synergistic Intelligence in Multimodal Large Language Models (MLLMs): Capability Boundaries, Hallucination Mitigation, and Reasoning Enhancement. Journal of Knowledge Engineering and Social Development, 3(1), 1–5. Retrieved from https://jkesd.com/index.php/JKESD/article/view/43
Issue
Section
Articles