Can large language models (LLMs) truly grasp mathematical reasoning, or are they just mimicking human calculations? A new research paper explores this question, venturing beyond the typical "Chain of Thought" (CoT) approach and delving into the world of Prolog, a logic programming language. CoT prompting encourages LLMs to generate step-by-step reasoning, but this can lead to cascading errors. Imagine a student meticulously outlining their math solution, only to stumble on a simple addition in the middle—the entire answer becomes wrong. This new research suggests that LLMs might be better off focusing on extracting the core 'facts' of a problem and formulating them into symbolic logic, letting an external tool handle the actual computation. Think of it like a detective gathering clues and presenting them to a forensic expert for analysis. The researchers used Prolog, a language built on logical predicates, to represent math problems. They found that LLMs generating Prolog code outperformed those using CoT on a standard math benchmark (GSM8K), especially when dealing with large numbers. This suggests that LLMs can effectively translate math problems into logical statements, even if they struggle with the calculations themselves. Furthermore, the researchers introduced a novel technique called 'predicate permutation.' Since the order of facts in Prolog doesn't affect the outcome, they shuffled the order during training, forcing the LLM to learn the underlying logic more robustly. This is like teaching a student to solve a puzzle from different starting points, strengthening their understanding of the overall picture. While this research shows promise, challenges remain. The current Prolog interpreter has limitations, and the impact of model size on this approach is still unknown. However, this work opens exciting avenues for integrating symbolic reasoning with LLMs, potentially leading to more reliable and transparent AI problem-solving in the future.