Multimodal Visual Question Answering (VQA) has received significant attention as it has the capability to solve intricate problems, such as puzzles, by relating visual and text information. This paper is centered on enhancing the algorithmic reasoning process of Large Language Models (LLMs) for such activities. In this paper Vision Transformers (ViT base / DeiT base) are employed to extract visual features, attention mechanisms such as cross attention and scaled dot product attention are employed for better multimodal fusion. DeiT base with cross and scaled dot product attention mechanism resulted in mean accuracy of 61.4% and 56.5% respectively, which is better than the existing LLMs. They also achieved maximum accuracy of 100% in puzzles such as Move Box, Rotten Fruits, Tower of Hanoi, Water Jug, and Wood Slide. The proposed model performed best in 10 of 18 puzzle categories compared to models like GPT4 Turbo, GPT-4o and GPT o1. Using vision transformer (ViT) the image feature extraction ability is enhanced and finding relation between image feature embeddings, textual embeddings is improved by using cross attention mechanism. AlgoPuzzleVQA dataset was used in this paper that consist of 18 different algorithmic puzzles resulting in 1800 instances in total. The new multimodal VQA model is set to find various real-world applications, including aiding visually impaired people, enhancing medical diagnosis, and providing AI-supported educational guidance.