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

Problem Statement

As Large Language Models (LLMs) are increasingly deployed in decision-support workflows ranging from medical diagnosis to financial planning, understanding their behavior under uncertainty has become critical. While classical decision theory often models agents as rational utility maximizers, empirical evidence in human contexts shows that decision-making is shaped by stable dispositional risk preferences and context-dependent adjustments. It remains unclear whether LLMs exhibit analogous, stable, and interpretable behavioral signatures of risk-sensitive decision-making, or if their choices are merely random variations around a common policy. Specifically, the paper investigates two core questions:

1. Trait Component: Do different LLMs exhibit stable, model-specific risk profiles (e.g., conservative vs. aggressive) that persist across repeated interactions and changing social environments?
2. Adaptive Component: Do LLMs regulate their behavior in response to changing risk conditions, such as increased global risk pressure or personal resource constraints?

Methodology

The authors employed a controlled multi-model framework based on No-Limit Texas Hold'em, a stochastic game characterized by incomplete information and sequential betting. This environment serves as a proxy for real-world decision-making, requiring agents to balance potential gains against the risk of loss.

Experimental Design

The study evaluated six frontier LLMs: GPT-5.4, Claude-Sonnet-4.6, Gemini-3.1-Pro, DeepSeek-V4-Pro, Xiaomi MiMo-V2.5-Pro, and Qwen3.6-Plus. The experiments were structured to isolate behavioral traits from environmental noise:

1. Homogeneous Self-Play: All six table positions were occupied by instances of the same model. This setup isolated intrinsic model-specific behavioral priors from opponent effects.
2. Heterogeneous Mixed-Model Interactions: Six different models competed at the same table to test the robustness of individual risk profiles under social pressure and varying opponent dispositions.
3. Risk Manipulations:
   • Global Risk Pressure: The blind levels were increased (from 10bb to 50bb, 100bb, and 200bb) while keeping initial stacks fixed, increasing the cost of participation for all agents.
   • Personal Risk Exposure: The stack size of a focal model was reduced (to 500 or 100 chips) while the surrounding environment remained unchanged, testing responses to asymmetric resource constraints.

Behavioral Metrics

Behavior was quantified using two interpretable pre-flop metrics:

• Participation (VPIP): The frequency of voluntarily entering the pot (excluding forced blinds), measuring a model's intrinsic willingness to engage in uncertain opportunities.
• Proactiveness (PFR): The frequency of making pre-flop raises, measuring a model's propensity for risk-escalating, assertive actions.

These metrics defined a behavioral state space to map risk profiles from conservative (low VPIP, low PFR) to aggressive (high VPIP, high PFR).

Key Results

1. Stable Trait-Level Risk Profiles

In homogeneous self-play, models occupied distinct, non-overlapping regions in the Participation–Proactiveness space, forming a spectrum from conservative to aggressive:

• Conservative Group: Gemini, Qwen, and DeepSeek exhibited low participation (18–23%) and low proactiveness (6–15%).
• Moderate Group: Claude and Xiaomi showed intermediate profiles.
• Aggressive Group: GPT-5.4 demonstrated the highest participation (45%) and proactiveness (21%).
  These differences were larger than within-model fluctuations, indicating stable, model-specific behavioral priors rather than random noise.

2. Social Robustness and Divergence

When placed in heterogeneous mixed-model settings:

• Robustness: Most models (Qwen, DeepSeek, Claude, Xiaomi) preserved their baseline risk profiles, showing no statistically significant shifts in participation or proactiveness.
• Divergence: The extremes amplified. The most aggressive model (GPT-5.4) became significantly more aggressive, while the most conservative model (Gemini) became even more cautious. This suggests a form of social polarization where mixed interactions do not induce convergence to a shared norm but rather accentuate existing dispositions.

3. Impact on Competitive Outcomes

Behavioral profiles translated into distinct outcome structures:

• In homogeneous play, all models achieved similar overall win rates (~17%), but via different strategic routes (e.g., conservative models had higher win rates on entered hands but lower volume; aggressive models had lower win rates on entered hands but larger average win stacks).
• In heterogeneous play, these differences led to unequal competitive returns. GPT-5.4 achieved the highest overall win rate (30.77%), while conservative models saw their win rates drop significantly (e.g., Gemini to 8.13%), indicating that aggressive engagement can be advantageous in mixed environments.

4. Context-Dependent Adaptation

Models exhibited heterogeneous responses to risk changes:

• Global Risk Pressure: Increasing blind levels generally reduced risk-taking, but adaptation patterns varied:
  • Broad Contraction: Claude, Xiaomi, and DeepSeek reduced both participation and proactiveness significantly.
  • Selective De-escalation: GPT-5.4 maintained participation but significantly reduced proactiveness (raising less often once involved).
  • Near-Invariant: Gemini showed minimal change, preserving its baseline policy even under high global pressure.
• Personal Risk Exposure: When a focal model's stack was reduced, most models became more cautious (reducing participation), but again in model-specific ways. GPT-5.4 maintained high participation but reduced aggression, while Gemini remained largely insensitive to its own resource constraints.

Significance and Claims

The paper claims to provide a behavioral science perspective on LLMs, demonstrating that they are not interchangeable decision engines under uncertainty. The key contributions are:

1. Identification of Behavioral Signatures: LLMs exhibit stable, distinguishable risk profiles (traits) that are identifiable across repeated interactions and social contexts.
2. Differentiation of Adaptation: Models differ not only in their baseline risk disposition but also in their risk plasticity—how they respond to global risk signals versus personal resource constraints.
3. Audit Framework: The study proposes that evaluating AI-assisted decision-making requires looking beyond task performance or reasoning accuracy. Instead, systems should be audited for their trait-level disposition, social robustness, global risk plasticity, and personal resource sensitivity.

The authors conclude that these findings suggest LLMs differ in the risk signals they respond to and the flexibility with which they adjust. This provides a behavioral basis for auditing risk-sensitive decision-making in interactive settings, implying that future evaluations should report comprehensive behavioral risk profiles to ensure models are matched appropriately to the risk tolerance of specific tasks (e.g., conservative models for safety-critical settings vs. proactive models for exploratory environments).

The paper does not claim that LLMs possess human-like psychological risk preferences or cognitive mechanisms; rather, it argues that their observable decisions can be organized using behavioral constructs useful for studying decision-making under uncertainty.