Do Language Models Exhibit the Same Cognitive Biases in Problem Solving as Human Learners?

TL;DR

This study investigates whether large language models exhibit similar cognitive biases as children when solving arithmetic word problems, revealing that LLMs share biases in comprehension and planning but not in execution.

Contribution

The paper introduces a novel testing framework and a neuro-symbolic approach to analyze LLMs' problem-solving steps and biases in comparison to human learners.

Findings

LLMs show human-like biases in text comprehension

LLMs exhibit similar biases in solution planning

LLMs do not show biases in arithmetic execution

Abstract

There is increasing interest in employing large language models (LLMs) as cognitive models. For such purposes, it is central to understand which properties of human cognition are well-modeled by LLMs, and which are not. In this work, we study the biases of LLMs in relation to those known in children when solving arithmetic word problems. Surveying the learning science literature, we posit that the problem-solving process can be split into three distinct steps: text comprehension, solution planning and solution execution. We construct tests for each one in order to understand whether current LLMs display the same cognitive biases as children in these steps. We generate a novel set of word problems for each of these tests, using a neuro-symbolic approach that enables fine-grained control over the problem features. We find evidence that LLMs, with and without instruction-tuning, exhibit human-like biases in both the text-comprehension and the solution-planning steps of the solving process, but not in the final step, in which the arithmetic expressions are executed to obtain the answer.