Scholarlink

<p>Large language models (LLMs) are increasingly used for decision support and behavioral simulation, yet it remains unclear whether they reproduce classic mental accounting biases and how these patterns depend on model configuration. This study examines whether LLMs exhibit hedonic framing, account non-fungibility, and sunk cost sensitivity, and how these tendencies respond to demographic personas and sampling temperature. We evaluate Gemini 2.0 Flash, Claude 3.5 Sonnet, GPT 4o, and DeepSeek V3 on a curated set of canonical decision scenarios that are also administered to 306 human participants in text-only tasks. Each item is replicated across multiple independent runs with majority aggregation, and we estimate prospect theory parameters and fit regression models to compare decision patterns across humans and models. The results show that LLMs often mirror human-like mental accounting, including segregating gains and losses, treating nominally equivalent resources as non-fungible, and honoring sunk costs, although the strength and direction of these effects differ across models and conditions. Demographic personas and temperature systematically modulate these biases, shifting model behavior between more rational and more human-biased regimes without eliminating mental accounting effects. These findings clarify when LLMs can approximate human decision behavior in economic tasks and provide practical guidance for configuring LLM-based agents in applications such as project continuation, pricing, and portfolio choices.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Do LLMs Keep Mental Accounts? Empirical Evidence from Hedonic Framing, Nonfungibility, and Sunk Cost Sensitivity

  • 期刊论文

  • 发表日期：

    2026年8月3日

查看全文
  • Zhen-Song Chen,
  • Wei-Chen Jiao,
  • Y. P. Tsang,
  • C. H. Wu,
  • Xian-Jia Wang

摘要

Large language models (LLMs) are increasingly used for decision support and behavioral simulation, yet it remains unclear whether they reproduce classic mental accounting biases and how these patterns depend on model configuration. This study examines whether LLMs exhibit hedonic framing, account non-fungibility, and sunk cost sensitivity, and how these tendencies respond to demographic personas and sampling temperature. We evaluate Gemini 2.0 Flash, Claude 3.5 Sonnet, GPT 4o, and DeepSeek V3 on a curated set of canonical decision scenarios that are also administered to 306 human participants in text-only tasks. Each item is replicated across multiple independent runs with majority aggregation, and we estimate prospect theory parameters and fit regression models to compare decision patterns across humans and models. The results show that LLMs often mirror human-like mental accounting, including segregating gains and losses, treating nominally equivalent resources as non-fungible, and honoring sunk costs, although the strength and direction of these effects differ across models and conditions. Demographic personas and temperature systematically modulate these biases, shifting model behavior between more rational and more human-biased regimes without eliminating mental accounting effects. These findings clarify when LLMs can approximate human decision behavior in economic tasks and provide practical guidance for configuring LLM-based agents in applications such as project continuation, pricing, and portfolio choices.