Title: Examining the Gap Between Outcome Similarity and Mechanism Consistency in LLM Risk Assessment: Insights from the St. Petersburg Paradox

Abstract:

While large language models (LLMs) may exhibit caution in risk-related tasks, outwardly prudent outputs do not necessarily reflect alignment with human decision-making processes. To explore this distinction, we utilized the St. Petersburg game—a classic paradox characterized by infinite expected value but low human willingness to pay—as a controlled experimental framework. Our study assessed 28 LLMs using a comprehensive prompt suite that featured the original game, as well as controlled variants manipulating factors such as truncation, repeated play scenarios, numeric endowments, and occupational identity. The suite also included prompts designed to elicit human-perspective reasoning and paired comparisons between base models and those subjected to instruction tuning.

In the baseline St. Petersburg task, the majority of models produced finite bids, superficially mimicking human risk aversion. However, this outcome-level similarity obscures significant divergences in underlying mechanisms. Analysis of the controlled variants indicates that models frequently abandon human-like behavior in favor of conditional and computational rationality when specific constraints are introduced. Although instruction tuning and human-cue prompting generally resulted in lower bids and mitigated certain observable anomalies, the fundamental mechanism-level response patterns remained largely stable. These results suggest that behavioral alignment in risk decision-making can be merely superficial; LLMs can generate human-like decisions without adopting human-consistent reasoning processes. Consequently, high-stakes evaluations of LLM decision-making must transcend simple outcome similarity to rigorously assess mechanism-level consistency.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC