Summary of https://arxiv.org/pdf/2410.05229 This research paper investigates the mathematical reasoning capabilities of large language models (LLMs) and finds that their performance is not as robust as previously thought. The authors introduce a new benchmark called GSM-Symbolic, which generates variations of math problems to assess the models' ability to generalize and handle changes in question structure. The results show that LLMs struggle to perform true logical reasoning, often exhibiting a high degree of sensitivity to minor changes in input. The authors also find that LLMs often blindly follow irrelevant information in the questions, suggesting that their reasoning process is more like pattern matching than true conceptual understanding.