Study Reveals Two Distinct LLM Categories in Risky Decision-Making

A research paper published on arXiv examines how large language models handle risky choices, identifying two distinct behavioral clusters. The study involved 20 frontier and open LLMs, analyzing their decision-making under uncertainty across two dimensions: prospect representation and decision rationale. Reasoning models (RMs) demonstrated rational behavior, showing insensitivity to prospect order, gain/loss framing, and explanations, while conversational models (CMs) displayed significant differences in these areas. The research was complemented by a matched human subjects experiment, providing a human reference point, and compared against an expected payoff maximizing rational agent model. The paper, arXiv:2602.15173v2, explores how LLMs function as decision support systems or in agentic workflows within the digital ecosystem. Findings indicate that understanding of LLM decision-making under uncertainty remains limited despite rapid transformation of digital environments. The study's methodology included examining whether prospects were presented explicitly or through outcome histories. This research contributes to understanding how different LLM architectures approach risk assessment and decision justification.