This paper investigates the interpretability of reasoning explanations generated by prompting large language models (LLMs). The authors evaluate the faithfulness, robustness, and utility of reasoning chains produced by various prompting techniques, including Chain-of-Thought (CoT), Self-Consistency (SC-CoT), and their proposed Self-Entailment-Alignment CoT (SEA-CoT). The study shows that SEA-CoT significantly improves interpretability across multiple dimensions, achieving over 70% improvements in several metrics. The authors also conduct extensive experiments on three commonsense reasoning benchmarks: OpenBookQA, QASC, and StrategyQA. They find that SEA-CoT outperforms other methods in terms of faithfulness, robustness, and utility. The study highlights the importance of aligning explanations with the context and the intended answer to enhance interpretability. The authors propose a simple alignment technique that enhances the quality of explanations by ensuring they are consistent with the context and the final answer. The results demonstrate that SEA-CoT provides more interpretable explanations than other methods, making it a promising approach for improving the transparency and explainability of LLMs. The study also discusses the limitations of current methods and suggests future research directions in the area of interpretable AI.This paper investigates the interpretability of reasoning explanations generated by prompting large language models (LLMs). The authors evaluate the faithfulness, robustness, and utility of reasoning chains produced by various prompting techniques, including Chain-of-Thought (CoT), Self-Consistency (SC-CoT), and their proposed Self-Entailment-Alignment CoT (SEA-CoT). The study shows that SEA-CoT significantly improves interpretability across multiple dimensions, achieving over 70% improvements in several metrics. The authors also conduct extensive experiments on three commonsense reasoning benchmarks: OpenBookQA, QASC, and StrategyQA. They find that SEA-CoT outperforms other methods in terms of faithfulness, robustness, and utility. The study highlights the importance of aligning explanations with the context and the intended answer to enhance interpretability. The authors propose a simple alignment technique that enhances the quality of explanations by ensuring they are consistent with the context and the final answer. The results demonstrate that SEA-CoT provides more interpretable explanations than other methods, making it a promising approach for improving the transparency and explainability of LLMs. The study also discusses the limitations of current methods and suggests future research directions in the area of interpretable AI.