LLM economicus? Mapping the Behavioral Biases of LLMs via Utility Theory

Thank you for your thoughtful review. We are glad you found our methods rigorous, comprehensive, and well-communicated. We likewise believe that this work presents a valuable set of tools for designing LLMs for economic decision-making, and we will be releasing an open-source package to ensure that others can easily adopt these tools.

Question 1 (competency tests). Thank you for your question, and we are happy to provide further rationale for the competence test for each game presented in the paper. For the ultimatum game, some LLMs did not pass the competence test because they were unable to do basic arithmetic (i.e. 10 dollars - 6 dollars = 4 dollars). For the gambling game, some LLMs did not pass the competence test because they were logically inconsistent about the gambles they take. For example, an LLM would be incompetent if the LLM was willing to bet 50 dollars at 20% probability, unwilling to bet 60 dollars at 20% probability, and willing to bet 70 dollars at 20% probability. For the waiting game, some LLMs did not pass the competence test because they were logically inconsistent about how long they were willing to wait for a monetary gain. For example, an LLM would be incompetent if the LLM was willing to wait 5 years for 50 dollars, unwilling to wait 10 years for 50 dollars, but willing to wait 25 years for 50 dollars. We will provide more examples of these failure modes and discuss this further in the next iteration of the paper.

Thank you for your thoughtful review.

Weakness 1 (experimental lightness). We would like to push back on this point, as we performed prompt ablations (Appendix B.3, C.3, D.3) on the system and user prompts. For most ablations, we found that the formulation of the prompts had little effect on the strategic choice given repeated samples of the learned distribution. We attempted as much experimental rigor as possible within a limited research budget of compute and API credits. We will make this point clearer in the next iteration of the paper.

Weakness 2 (competence tests). The competence tests are designed to test specific competencies related to playing each game. We assume that we cannot meaningfully analyze an LLM’s behavior if the LLM cannot execute the basic reasoning capabilities to play the game. (We also tried various prompt ablations to ensure that the competence tests were robust). Indeed, our tests are meaningful only if we (at least weakly) assume that LLMs “are understanding these texts like people do”. While this is a valid critique, this assumption is implicit in pretty much every prompt-based interpretability work, which is one of the only means by which we can probe the most capable LLMs.

Weakness 3 (presentation). We will explicitly explain variables in the accompanying captions for each figure and update plots to be differentiated by pattern in addition to color.

Weakness 4 (human results). Human results are taken from (classic) studies in experimental economics; we did not run our own human studies. We report parameters from papers cited in each game (Fehr & Schmidt 1999, Kahneman & Tversky 1992, Rachlin et al 1991 respectively); we will clarify how we compare human behavior in Section 3 and explicitly cite each value in figure footnotes. We believe the human results are comparable because we directly adapt the experiments reported in these papers to LLMs.

Weakness 5 (personas). Section 4 was intended as a case study, and we intend to further examine the effect of personas and role-playing in follow-on work. The main contribution of our paper is to show that utility theory provides interesting tools for probing the economic behavior of LLMs. Whether and how LLM behavior can be changed through prompting (e.g., roles-based prompting) or finetuning is an interesting avenue for future work.

Weakness 6 (language choice). “Understand” and the ELIZA anecdote has unintended implications and will update accordingly.

Thank you for your thoughtful review.

Weakness 1 (technical depth). We present a novel set of benchmarks and evaluation protocols - a topic of interest for COLM. We claim that our methods enable further studies into what impacts economic decision-making abilities, as shown in Section 4. For example, we show how our evaluation protocols can be used to study how implicit assumptions like age group affect an LLM’s risk and loss aversion, enabling researchers to probe for proxies that impact economic decisions. We focused on prompting techniques as the most economically interesting/competent models (i.e., GPT-4) were only accessible via prompting. However, we believe that the framework we present could be used to provide feedback to LLMs during fine-tuning, and such a technique could become practically relevant as more capable open-source models are released.

Weakness 2 (reasoning and decision-making). Mathematical reasoning is closely related to rationality of economic decision-making insofar as it provides a set of tools necessary to analyze and model economic decisions. For example, the decision for whether or not to take an economic gamble (on a stock, insurance plan, or new home) relies on mathematical reasoning to calculate the expected value over time. However, human economic decision making is not rational in many cases, and some of the economic experiments that we consider in the study were originally designed to surface this irrationality.

Thank you for your thoughtful review. We are glad you found our work novel and useful in comparing LLM behavior to human behavior. We believe that the ability to leverage existing human studies to quantifiably compare LLM behavior is a core strength of our approach.

Weakness 1 (premise). The reviewer suggests that because LLMs do not have grounding in the physical world, what they “say” may be different than what LLMs “do”, and hence probing the economic “behavior” of LLMs doesn’t really make sense. We agree that this is a valid (and deep!) point, and acknowledged some of these limitations in Section 5. However, we note that there is a rich body of work on treating LLMs as human-like agents in the context of economic/behavioral experiments (e.g., https://arxiv.org/abs/2208.10264 https://www.nber.org/papers/w31122 ), and these (influential) studies have yielded interesting findings and have spurred on much follow-up work. Moreover, we observe that even in the case of humans, in some of the experiments the questions are posed only as hypotheticals; we trust that the answers given by the experimental subjects reflect what they would truly do. From this point of view, we think that our experimental setup (as well as the experimental setup of the aforementioned studies) is not invalid. We nonetheless acknowledge that this point needs to be discussed more thoroughly, and will do so in the next iteration of the paper.

Weakness 2 (competence test). This is a good question. We do not see obvious patterns in the amount of bias as a function of their competence. To clarify further, the competence tests are designed to test specific competencies related to playing each game. For example, the competence test for the ultimatum game tests whether the LLM can do the requisite arithmetic to make an offer. We assume that we cannot meaningfully analyze an LLM’s behavior if the LLM cannot execute the basic reasoning capabilities to play the game. Given this, we agree that it may be hard to make general economic statements about LLMs as a class, or even as a single LLM family, given that an LLM may show competence in one experiment but not in another. However, once the LLM has passed the competence test, we can identify specific behavioral biases in specific versions of LLMs that may have profound impact on deploying these models in decision-making settings. We will add more discussion around this point.