Uncover the amazing secret of the big language model: LLM can understand space and time
Research from MIT shows that large language models (LLMs), with their ability to deeply understand space and time, may be more than a simple text-generating tool. Through experiments, the researchers found that the model was able to predict geographic location and time information around the world, including the year of death of historical figures, the date of publication of works of art, and the publication date of news headlines. In addition, the researchers analyzed the spatial and temporal representations inside the model and found that the model has the ability to map from place to country, and has a certain understanding of the global geographical structure of different groups in geospace (or time). These findings bring new thinking and possibilities for the development and application of language models in the future.