LLMs Exhibit Stable, Interpretable Risk-Sensitive Decision-Making.

As large language models (LLMs) are increasingly integrated into decision support systems, understanding their behavior under uncertainty is paramount. Human decision-making typically involves consistent risk preferences alongside context-dependent adjustments. This research explores whether similar stable and interpretable behavioral regularities can be observed in LLM-based decision systems. Using a controlled multi-model framework based on no-limit Texas Hold'em poker, researchers quantified LLM behavior through "Participation" (voluntary engagement in uncertain opportunities) and "Proactiveness" (pre-flop risk escalation). Across both homogeneous self-play and heterogeneous mixed-model interactions, frontier LLMs consistently exhibited stable, model-specific risk profiles, spanning a spectrum from conservative to aggressive decision styles. These risk profiles largely remained robust even when opponent compositions changed, with the most conservative and aggressive models diverging further in mixed settings. When subjected to global risk pressure and personal resource constraints, models adapted in structured but varied ways, including broad behavioral contraction, selective de-escalation, or near-invariant behavior. These findings suggest that LLMs differ not only in their baseline risk disposition but also in how they perceive and respond to risk signals, and their flexibility in adjusting behavior, offering a crucial foundation for auditing risk-sensitive decision-making in interactive AI systems.