#278 скачать в хорошем качестве

#278

Large Language Models (LLMs) are increasingly utilized in autonomous decision-making, where they sample options from vast action spaces. However, the heuristics that guide this sampling process remain under-explored. This study examines the sampling behavior and shows that the underlying heuristics resemble those of human decision-making: comprising a descriptive component (reflecting statistical norm) and a prescriptive component (implicit ideal encoded in the LLM) of a concept. It is demonstrated that the deviation of a sample from the statistical norm towards a prescriptive component consistently appears in concepts across diverse real-world domains like public health and economic trends. To further illustrate the theory, the study shows that concept prototypes in LLMs are affected by prescriptive norms, similar to the concept of normality in humans. Through case studies and comparison with human studies, it is illustrated that in real-world applications, the shift of samples toward an ideal value in LLMs’ outputs can result in significantly biased decision-making, raising ethical concerns. In this video, I talk about the following: Are LLMs biased towards making descriptive choices or prescriptive choices? How about the bias in LLM choices for existing concepts? How much do LLM choices correlate with humans? For more details, please look at https://aclanthology.org/2025.acl-lon... Sivaprasad, Sarath, Pramod Kaushik, Sahar Abdelnabi, and Mario Fritz. "A theory of response sampling in LLMs: Part descriptive and part prescriptive." In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pp. 30091-30135. 2025. Thanks for watching! LinkedIn: http://aka.ms/manishgupta HomePage: https://sites.google.com/view/manishg/