Behavioural Signatures of Risk-Sensitive Decision-Making in Large Language Models

This paper studies whether large language models display stable and interpretable risk-sensitive decision patterns. Using no-limit Texas Hold’em, the authors evaluate multiple frontier models through Participation, which captures voluntary engagement in uncertain opportunities, and Proactiveness, which captures pre-flop risk escalation. In homogeneous self-play and heterogeneous mixed-model games, models exhibit distinct conservative-to-aggressive profiles. These profiles are generally robust to opponent composition, although extreme models diverge more in mixed settings. Under global risk pressure and personal resource constraints, models show heterogeneous adaptations, including broad contraction, selective de-escalation, and near-invariant behaviour. The AgentTexasPoker code is publicly available.