Will the Real Linda Please Stand up...to Large Language Models? Examining the Representativeness Heuristic in LLMs

Dear reviewer 4fxrR,

Thank you for your constructive comments and for recognizing the overall value of our work.

• Change Subtitle 5.3: “Improving Performance by Hinting (that LLMs to ?) use Their Knowledge”

We have changed Subtitle 5.3 accordingly to “Improving Performance by Hinting that LLMs Use Their Knowledge.”

• Discussing your results in light of human base rates would be interesting.

After receiving IRB approval from our institution, we have completed human evaluations on our proposed ReHeAT dataset. Through the Prolific platform, we recruited 153 participants to answer questions. Participants are asked to answer 3-4 types of representative heuristic questions. Human results are close to those of LLaMA2-70B, whereas the GPT series performs better than the human results. We will incorporate these new results into our final version and provide more comprehensive outcomes.

• I see the note on page 8 that human base rates are around 10 to 50%. In a way, should we expect LLMs to exhibit this bias, in that they are trained to replicate the judgments of a population of humans?

Thank you for asking this great question. We do observe similar performance between humans and LMs. However, regarding whether we should expect LLMs to exhibit the representativeness heuristic, we think it depends on the LMs’ potential use. When LLMs are used to aid in decision-making, we certainly do not want them to display such biases, as this cognitive trap can lead to incorrect judgments. From another viewpoint, this heuristic is a simplification strategy that helps people make quick decisions, which can be beneficial. If our goal is for LLMs to mimic human behavior closely, we should expect them to exhibit the representativeness heuristic. In reality, however, our aim extends beyond merely replicating human-like behavior; we also hope that LLMs will surpass human performance in areas prone to bias and other forms of error.

Dear reviewer f6wA,

Thank you for your positive and constructive feedback. Below, we respond to your questions.

• Do you think more sophistic prompting strategies such as multi-agent debating can bring larger improvements?

We believe more sophisticated prompting strategies could bring improvements, similar to how humans, when faced with heuristic questions, think deeply, and invest more mental resources to avoid heuristic biases. We will explore more complex prompting strategies in our future work.

• It would be better to provide experiments of testing models performance across different scales.

Thank you for your advice. We provide additional experiments on LLaMA2-13B to compare with LLaMA2-70B. The performance of LLaMA2-13B is slightly lower than that of LLaMA2-70B. We can infer that model size impacts performance on representativeness heuristic questions, with larger models generally enhancing reasoning capabilities to a certain extent.

• Is there any results on average human accuracy on the proposed testset?

After receiving IRB approval from our institution, we have completed human evaluations on our proposed ReHeAT dataset. Through the Prolific platform, we recruited 153 participants to answer questions. Participants are asked to answer 3-4 types of representative heuristic questions. Human results are close to those of LLaMA2-70B, whereas the GPT series performs better than the human results. We will incorporate these new results into our final version and provide a more comprehensive analysis.

Note. Here we compare Human results with the LMs’ Standard results of all other models.

Dear reviewer eSG3,

We sincerely thank your insightful comments and valuable suggestions. Below, we address the important concerns you raised:

• It would be helpful if the authors could conduct similar human studies.

After receiving IRB approval from our institution, we have completed human evaluations on our proposed ReHeAT dataset. Through the Prolific platform, we recruited 153 participants to answer questions. Participants are asked to answer 3-4 types of representative heuristic questions. We will incorporate these new results into our final version and provide more comprehensive outcomes.

• The curation process for the new data is unclear. How diverse are the new questions?

The 153 new questions were organized to retain the essence of the representativeness heuristic being tested, at the same time changing the situations and contexts of the original items. Because each type of representativeness heuristic question shares the same statistical prototype, we aimed to display diversity in the contexts of the questions. There are approximately 15 to 18 different contexts for each type of question.

• It would be better if the authors could include results using more examples for SPQs and IQs.

Thank you for this advice. Each type of representativeness heuristic has the same statistical prototype. Therefore, only one SPQ is needed for each type, and we collected the majority of answers for each SPQ through self-consistency. Although IQs could indeed benefit from more examples for demonstration, considering that one of the purposes of IQs is to directly compare with SPQs, we adopted the same test format as the SPQs.

• How are the ICL examples selected in section 5.1?

The examples in Section 5.1 for IC are randomly selected from diverse backgrounds.

• Missing reference that compares human and LLMs reasoning, e.g. [1]

Thank you for recommending the missing references. We will add this reference in our final version.

Dear reviewer 3FvA,

Thank you for your encouraging, thorough, and positive feedback. Below, we address your comments.

• What would really improve this paper a lot is spending some time discussing construct validity and external validity. Why are we testing models for these kinds of biases? What do we conclude about them when they get these results?

We agree that further discussing construct validity and external validity will improve this paper, and we will add further discussion to the final version. To address construct validity, psychologists develop measures and apply psychometrics to measure individual differences in personal characteristics (e.g., traits, behaviors) as well as individual differences in intra-individual change (e.g., states, behavioral patterns). In a similar spirit of construct validity, here seek to measure the consistency between models, so that any empirical convergence provides indirect evidence of construct validity. Additionally, in psychology, common methods for improving external validity include improving the diversity of samples and improving the psychological fidelity of the data, so that both are more likely to generalize across situations. In this vein, when we explicitly attempt to provide diverse contexts for each type of representativeness heuristic and more realistic descriptions. We test the model's performance on these diverse and high-fidelity questions in our dataset to achieve external validity.

• The base rate fallacy picture doesn’t seem to be a great example of the fallacy. Would be nice if you could illustrate it there too. Unless I’m missing something?

Thank you for the advice regarding the base rate fallacy figure. We agree that although the current image reflects the Bayesian theorem of the base rate fallacy, it does not demonstrate its counterintuitive aspect. We will change the figure in the following way: (a) the blue dots within the large circle, but outside the small circle, will change to red, and (b) the red dots within the large circle, but outside the small circle, will change to blue. Revising in this manner shows how a high proportion of those who are sick have symptoms, but that does not mean that a high proportion of those with symptoms are sick.

• The caption on the Disjunction Fallacy image is ungrammatical.

We will change it to “A or B. A. People may think if A, then not B. However, if A is a subset of B, this conclusion would be incorrect.”