Multimodal Reasoning in Large Language Models: Cognitive Abilities, Limitations, and Open Challenges

Multimodal Large Language Models (MLLMs) have emerged as a transformative paradigm in artificial intelligence, extending traditional unimodal LLMs by jointly processing text, images, video, and other sensory modalities. While these models achieve impressive performance on perception-centric tasks, their ability to perform true reasoning across modalities including compositional, temporal, causal, and abstract inference remains limited. This survey provides a comprehensive, reasoning-centric overview of MLLMS, introducing a taxonomy of multimodal reasoning abilities, analyzing cognitive and architectural limitations, and identifying key research gaps. We examine current mechanisms that enable multimodal reasoning, highlight prevalent failure modes such as hallucinations and spurious correlations, and discuss challenges in data, model design, and evaluation. By synthesizing insights from recent high-impact studies, this work outlines future directions for developing interpretable, generalizable, and human-aligned multimodal reasoning systems, laying a foundation for progress toward artificial general intelligence.

کلیدواژه ها:

Multimodal Large Language Models ، Multimodal Reasoning ، Cognitive Limitations ، Causal Inference ، Compositional Reasoning ، Artificial General Intelligence