Can AI truly reason like us? Recent advances in large language models (LLMs) have shown impressive progress in complex reasoning tasks, thanks to techniques like chain-of-thought (CoT) prompting. However, a new study reveals that these AI models often take shortcuts, arriving at correct answers through flawed logic. Researchers discovered that LLMs sometimes generate seemingly reasonable explanations (CoTs) that don't actually justify their conclusions. They identified two reasoning styles: "centralized," where the AI focuses on the last step of its reasoning, and "distributed," where it draws on multiple steps. The study found that the distributed style is more prone to these logical inconsistencies. Digging deeper, the researchers found that even when an LLM's explanation misses crucial information from the given context, the model can sometimes "recall" that missing information when it finally generates the answer. This disconnect between the explanation and the answer is what leads to the unfaithful reasoning. To address this, the researchers developed a method called "inferential bridging." This technique helps the LLM focus on the right information by providing additional hints during the reasoning process. It also filters out noisy or irrelevant explanations, leading to more accurate and logically sound reasoning. The results are promising, showing significant improvements in the faithfulness of LLM reasoning across various logical tasks. While this research focuses on open-source LLMs, it highlights a critical challenge in AI development: ensuring that AI not only gets the right answers but also arrives at them through valid reasoning. This is a crucial step towards building truly trustworthy and reliable AI systems.