```html Multimodal Large Language Models: Thinking in Spaces Multimodal Large Language Models (MLLMs) have made impressive progress in recent years in processing and interpreting various data types, including text, images, and videos. A central question that arises is to what extent these models are capable of extracting and processing spatial information from visual data, similar to how humans do. A recently published paper titled "Thinking in Space: How Multimodal Large Language Models See, Remember, and Recall Spaces" investigates precisely this capability of MLLMs. The study presents a novel benchmark called VSI-Bench (Video-based Spatial Intelligence Benchmark), specifically designed to evaluate the visual-spatial intelligence of MLLMs. VSI-Bench consists of over 5,000 question-answer pairs based on video sequences. The questions test the models' spatial understanding by asking, for example, about the position of objects in space, spatial relationships between objects, or navigation within an environment. The results of the study show that while MLLMs exhibit some visual-spatial intelligence, it still lags significantly behind human capabilities. The models perform less well on tasks that require more complex spatial reasoning. This suggests that the spatial reasoning abilities of MLLMs still need further improvement. To understand how MLLMs process spatial information, the researchers analyzed the models both linguistically and visually. It was revealed that the models are capable of developing local world models and a certain degree of spatial awareness. This means they are able to construct a simplified representation of the environment and capture the position of objects within this representation. Interestingly, common linguistic reasoning techniques such as Chain-of-Thought, Self-Consistency, and Tree-of-Thoughts proved ineffective in improving the performance of MLLMs in the VSI-Bench. In contrast, the explicit generation of cognitive maps during the question-answering process led to an improvement in the models' ability to estimate spatial distances. This suggests that integrating explicit spatial representations into MLLMs could be a promising approach for improving their visual-spatial intelligence. The development of MLLMs with improved spatial abilities is relevant for a variety of applications, including robotics, navigation, autonomous driving, and the development of intelligent assistants. The ability to extract and process spatial information from visual data is crucial for a comprehensive understanding of the world and for interacting with it. The results of this study provide important insights into how MLLMs "think in spaces" and what challenges remain to be overcome to bring their visual-spatial intelligence to a level comparable to that of humans. Research in this area is of great importance for the future development of AI systems capable of solving complex tasks in the real world. Bibliography Yang, J., Yang, S., Gupta, A. W., Han, R., Fei-Fei, L., & Xie, S. (2024). Thinking in Space: How Multimodal Large Language Models See, Remember, and Recall Spaces. *arXiv preprint arXiv:2412.14171*. BradyFU/Awesome-Multimodal-Large-Language-Models. (n.d.). *GitHub*. Retrieved from https://github.com/BradyFU/Awesome-Multimodal-Large-Language-Models Thinking in Space: How Multimodal Large Language Models See, Remember, and Recall Spaces. (n.d.). *PaperReading*. Retrieved from https://paperreading.club/page?id=273838 Lin, J., Ye, S., & Lau, R. W. H. (2024). Do Multimodal Large Language Models See Like Humans?. *arXiv preprint arXiv:2412.09603*. Gupta, A. W., Yang, J., Yang, S., Han, R., Fei-Fei, L., & Xie, S. (2025). *Taking the Next Step with Generative Artificial Intelligence: The Transformative Role of Multimodal Large Language Models in Science Education*. Hämäläinen, P., Tavast, M., & Kunnari, A. (2023, April). Evaluating large language models in generating synthetic hci research data: A case study. In *Proceedings of the 2023 CHI conference on human factors in computing systems* (pp. 1-19). Yin, S., Fu, C., Zhao, S., Li, K., Sun, X., Xu, T., & Chen, E. (2024). A survey on multimodal large language models. *National Science Review*, *11*(12), nwae403. louthy. (2024). *Hallucination is inevitable: An innate limitation of large language models (arxiv.org)*. Hacker News. Retrieved from https://news.ycombinator.com/item?id=39499207 Koh, P. W., & Liang, P. (2023, July). Foundation models for decision making: Problems, methods, and opportunities. In *ICML 2023 Workshop on Foundation Models for Decision Making* Yangyi-Chen/Multimodal-AND-Large-Language-Models. (n.d.). *GitHub*. Retrieved from https://github.com/Yangyi-Chen/Multimodal-AND-Large-Language-Models ```