From Strategic Narratives to Code-Like Cognitive Models: An LLM-Based Approach in A Sorting Task

We thank the reviewer for the insightful observations and valuable feedback on our work. We share many of your concerns about this work.

Comparing generated code with ground-truth algorithms: We compared participant behavior predictions generated by GPT-4 with ground-truth algorithms from RNNs. The generated codes performed worse (mean accuracy 30.5% vs. 40.4%), but GPT-4 performed better for more subjects. This suggests that GPT-4-generated codes may worsen predictions or that RNN algorithms capture some biases more accurately.

We used GPT-4 and LLaMA-3-70b to attribute a standard algorithm from 16 given algorithms, then used this prediction as standard approximated codes. Changing models or tasks did not improve accuracy compared to the original GPT-4 codes, directing the idea that RNN-attributed algorithms capture biases better.

Data inspection showed participants better predicted by RNNs primarily used the 'selection sort' strategy, involving simple heuristic steps. GPT-4 was sometimes misled by subjects’ descriptions, causing errors. For some, descriptions didn't match actions, leading to incorrect code generation. In other cases, GPT-4 implemented 'selection sort' more efficiently, causing errors. Simple heuristic descriptions, like 'swap each pair of images', also confused GPT-4.

These findings indicate our prompts need refinement. We'll use advanced prompting techniques to better guide LLM agents in solving this task.

Fundamental utility and limitations: We acknowledge that converting cognitive experiences from natural language or behavior into code requires significant abstraction. Currently, we don't expect LLMs to fully automate scientific discovery but to aid in generating hypotheses through code-like cognitive models. Traditional cognitive modeling involves iteratively fitting models to data, which can lead to overfitting and may miss important processes. LLM-generated codes expand the hypothesis space for cognitive models, serving as an initial step and benchmark. While LLMs can't capture all biases, they offer a human-understandable medium that aids model-building. This approach allows us to leverage LLMs' knowledge of sorting algorithms and human language to gain insights into human problem-solving strategies.

Thank you for your review. We appreciate your comments and have addressed the main concerns below.

1. Purpose of the Study The reviewer questioned the utility of using LLMs in our research. Our study investigates not the capability of humans and LLMs to solve the sorting problem, but their potential to reconstruct human cognitive processes in algorithm development based on human descriptions in naturalistic languages. This approach allows us to explore how LLMs leverage their inherent knowledge of sorting algorithms and humans’ language to simulate human problem-solving strategies, offering a unique intersection between AI capabilities and cognitive science insights.

2. Contributions to Cognitive Science The contribution of our research lies in demonstrating how LLMs can be utilized to model and understand human algorithmic thinking. By analyzing how LLMs reconstruct full algorithms from human descriptions, we gain insights into potential cognitive processes employed by humans, thus providing a novel method to explore human cognitive strategies through AI on a more detailed level.

3. Response to Specific Question We do not specify algorithm names; these algorithms are clustered in the previous study. In generations of codes, only human participants’ strategy descriptions are provided.

We thank the reviewer for the positive feedback on our paper. We also hope COLM could be a good place for this paper.

Addressing the gap between LLM and RNN: One approach to close the gap between LLM and RNN performance is to consider biases and random deviations from the described strategies. While RNNs easily handle these effects, the LLM-to-algorithm approach tends to ignore them unless explicitly described. A potential solution is for the LLM to output algorithms that generate likelihoods instead of hard choices. This would allow for a principled combination with other biases and suboptimalities inferred from choices. For example, randomness could be accommodated with a softmax parameter.

Prompt Engineering and Inductive Bias: We recognize that different prompts can introduce various inductive biases. This is a valuable consideration for future work to potentially close the gap between LLM and RNN performance.

Regarding the dataset and experimental design: It is important to note that our initial work used a preexisting dataset from an experiment not designed for this type of analysis. The original study ignored the verbal descriptions, leading to a limited number of trials per subject and after-the-fact strategy summaries. Moving forward, acquiring verbal data in real-time as participants complete the task, for example, via Think Aloud procedures, and including more trials and tasks per subject would enable more detailed analysis and better application of our approach.

We thank the reviewer for the positive feedback on our paper.

Regarding the use of different LLMs: Due to limited time and resources, we employed LLaMA-3-70b to address a similar task. We clustered human algorithms based on their descriptions and utilized the standard algorithm clustered by the LLM to predict human behavior. The results were comparably good (mean accuracy: 23.6%). The predictions made by different LLMs were consistent for most individuals (correlation = 0.55). Our findings indicate that using both GPT-4 and LLaMA-3-70b does not significantly improve accuracy, as over 75% of participants' accuracy differences were within 5%.

On the suggestion to try keywords and key phrases: This is an excellent idea. In our study, we instructed the LLM to generate code in a general format, adaptable to any sorting task length. Even when participants included action sequences in their descriptions, the LLM abstracted these sequences into executable and general codes. This abstraction suggests that the LLM can effectively distill the algorithms from the descriptions.

Concerning positional statements: As noted, backward algorithms were less accurately identified and predicted. Our supplementary analysis revealed that 6 of 8 standard algorithms exhibit a bias toward the forward version in predictive performance. However, in the recovery test (Fig. A2, appendix), backward algorithms were better identified with LLM self-generated strategy descriptions. We believe that this bias observed in empirical data likely stems from unequal sample sizes in human data and biases in human descriptions.

On individual differences in participant backgrounds: Little information is available beyond basic demographics, as participants were recruited online via Amazon’s Mechanical Turk. It is possible that the highest-performing participants had some education in sorting algorithms, providing clearer descriptions. This might also reflect the “good subject effect,” where participants who perform well on one task tend to perform well on all tasks, likely due to their higher investment in the experiment.