Decision-Making Behavior Evaluation Framework for LLMs under Uncertain Context

This research paper proposes a framework to assess the decision-making behaviors of large language models (LLMs) under uncertainty, finding that while LLMs generally exhibit human-like patterns of risk aversion and loss aversion, significant variations and potential biases emerge when socio-demographic features are introduced, highlighting the need for ethical guidelines in their deployment.

Abstract

When making decisions under uncertainty, individuals often deviate from rational behavior, which can be evaluated across three dimensions: risk preference, probability weighting, and loss aversion. Given the widespread use of large language models (LLMs) in decision-making processes, it is crucial to assess whether their behavior aligns with human norms and ethical expectations or exhibits potential biases. Several empirical studies have investigated the rationality and social behavior performance of LLMs, yet their internal decision-making tendencies and capabilities remain inadequately understood. This paper proposes a framework, grounded in behavioral economics, to evaluate the decision-making behaviors of LLMs. Through a multiple-choice-list experiment, we estimate the degree of risk preference, probability weighting, and loss aversion in a context-free setting for three commercial LLMs: ChatGPT-4.0-Turbo, Claude-3-Opus, and Gemini-1.0-pro. Our results reveal that LLMs generally exhibit patterns similar to humans, such as risk aversion and loss aversion, with a tendency to overweight small probabilities. However, there are significant variations in the degree to which these behaviors are expressed across different LLMs. We also explore their behavior when embedded with socio-demographic features, uncovering significant disparities. For instance, when modeled with attributes of sexual minority groups or physical disabilities, Claude-3-Opus displays increased risk aversion, leading to more conservative choices. These findings underscore the need for careful consideration of the ethical implications and potential biases in deploying LLMs in decision-making scenarios. Therefore, this study advocates for developing standards and guidelines to ensure that LLMs operate within ethical boundaries while enhancing their utility in complex decision-making environments.

Method

The authors used a multiple-choice-list experiment methodology grounded in behavioral economics theories to evaluate the decision-making behaviors of LLMs. This approach allowed them to estimate the degree of risk preference, probability weighting, and loss aversion in a context-free setting for three commercial LLMs: ChatGPT-4.0-Turbo, Claude-3-Opus, and Gemini-1.0-pro. They also explored the models' behavior when embedded with socio-demographic features of human beings, uncovering significant disparities across various demographic characteristics.

Main Finding

The authors discovered that LLMs generally exhibit decision-making behaviors akin to humans, such as risk aversion and loss aversion, with a tendency to overweight small probabilities. However, they found significant variations in the expression of these behaviors across different LLMs. Furthermore, when socio-demographic features were introduced into the decision-making processes, the LLMs displayed distinctive behavior patterns, with some models showing increased risk aversion in certain contexts and varying levels of risk aversion across different models. These findings suggest that LLMs may have inherent biases that need to be carefully considered when deploying them in decision-making scenarios.

Conclusion

The conclusion of the research is that while large language models (LLMs) generally exhibit decision-making behaviors similar to humans, with tendencies towards risk aversion and loss aversion, there are significant variations in these behaviors across different LLMs. The introduction of socio-demographic features into the decision-making processes of LLMs can lead to distinctive behavior patterns and potential biases. The authors advocate for the development of standards and ethical guidelines to ensure that LLMs operate fairly and ethically across diverse user groups and contexts.

Keywords

Large Language Models (LLMs), Decision-Making Behavior, Behavioral Economics, Risk Preference, Probability Weighting, Loss Aversion, Ethical Implications, Biases, Socio-Demographic Features, ChatGPT-4.0-Turbo, Claude-3-Opus, Gemini-1.0-pro, Context-Free Setting, Multiple-Choice-List Experiment, Evaluation Framework, Fairness, Ethical Decision-Making, Standards, Guidelines.