[2405.13012] Divergent Creativity in Humans and Large Language Models

PDF view of the paper entitled Creativity in Humans and Language Models, by Antoine Pelimary Beepin (1 and 2) and 39 other authors

PDF view

a summary:The recent increase in LLMS models led to allegations that they are close to the level of creativity, such as human capabilities. This idea sparked a mixture of excitement and fear. However, the critical piece that was missing in this discourse is a systematic evaluation of the semantic diversity of LLMS, especially compared to human disagreement thinking. To bridge this gap, we benefit from modern developments in mathematical creativity to analyze semantic difference in both modern LLMS and a large collection of data of 100,000 human beings. We have found evidence that LLMS can exceed the average human performance in the mission of the varying association, and approaches the capabilities of human creative writing, although it is not less than the typical performance of very creative humans. It is worth noting that even the best LLMS performance still exceeds largely by very creative individuals, confirming the ceiling that the current LLMS still fails to overcome it. It deals with the framework of the human measurement of the dialectical human machine surrounding the imminent replacement of human creativity by artificial intelligence, which leads to the lack of quality of creative linguistic outputs related to the use of fixed objective measures. As a deeper exploration of the distinctive elements of human creative thought compared to those of artificial intelligence systems, we put a series of techniques to improve their outputs regarding semantic diversity, such as rapid design and excessive parameter control.