Do Chains-of-Thoughts of Large Language Models Suffer from Hallucinations, Cognitive Biases, or Phobias in Bayesian Reasoning?
Roberto Araya
TL;DR
The paper investigates whether CoT reasoning in large language models naturally adopts ecologically valid strategies for Bayesian reasoning or exhibits biases toward symbolic reasoning. It uses a lie-detection Bayesian task with staged prompts to test three LLMs for adoption of natural frequencies, whole-object reasoning, and embodied heuristics. Results show a persistent bias toward probabilistic language with inconsistent uptake of the proposed strategies, suggesting an Einstellung-like resistance to embodied reasoning. These findings have implications for AI-assisted education, indicating a need for training and prompting approaches that incorporate novice-friendly reasoning and multimodal representations to improve pedagogical effectiveness.
Abstract
Learning to reason and carefully explain arguments is central to students' cognitive, mathematical, and computational thinking development. This is particularly challenging in problems under uncertainty and in Bayesian reasoning. With the new generation of large language models (LLMs) capable of reasoning using Chain-of-Thought (CoT), there is an excellent opportunity to learn with them as they explain their reasoning through a dialogue with their artificial internal voice. It is an engaging and excellent opportunity to learn Bayesian reasoning. Furthermore, given that different LLMs sometimes arrive at opposite solutions, CoT generates opportunities for deep learning by detailed comparisons of reasonings. However, unlike humans, we found that they do not autonomously explain using ecologically valid strategies like natural frequencies, whole objects, and embodied heuristics. This is unfortunate, as these strategies help humans avoid critical mistakes and have proven pedagogical value in Bayesian reasoning. In order to overcome these biases and aid understanding and learning, we included prompts that induce LLMs to use these strategies. We found that LLMs with CoT incorporate them but not consistently. They show persistent biases towards symbolic reasoning and avoidance or phobia of ecologically valid strategies.