Using Human Cumulative Prospect Theory to Understand Large Language Models Decision-Making

Abstract

Language models are trained to predict the next words for given content. The question arises: do large language models (LLMs), which have achieved technological breakthroughs in natural language generation, possess decision-making abilities as humans? The present article lets a series of LLMs do cognitive psychology experiments to explore this hypothesis. More specifically, we first identified a choice problem set and then conducted experiments following the certainty equivalent (CE) paradigm. We find that much of LLMs’ behaviors are consistent with humans: the fourfold pattern of risk attitudes can also be observed across the most LLMs subjects. Particularly, LLMs are risk-seeking and risk-averse for gains with low and high probabilities, LLMs are risk-averse and risk-seeking for losses with low and high probabilities. Furthermore, LLMs subjects exhibit greater rationality in losses than in gains, and less rationality during gains compared to losses. The exponential values of the value function and weighting function, fitted based on the data from LLMs subjects, satisfy the requirements of cumulative prospect theory (CPT). Taken together, these results will further enhance the theoretical foundations to understand the LLMs decision-making capabilities and provide novel tools from cognitive psychology for improving the interpretability of LLMs.

Cite this work

No public copy is stored for this record, so only the metadata is cited.

Stored only when both papers are in the corpus — 1 citations exist upstream.

Papers that cite the same references — the relational fallback while embeddings are still backfilling.