The paper "Towards Consistent Natural-Language Explanations via Explanation-Consistency Finetuning" addresses the issue of inconsistent explanations generated by large language models (LLMs). The authors propose a method called Explanation-Consistency Finetuning (EC-finetuning) to improve the consistency of LLMs' natural-language explanations across related examples. EC-finetuning involves finetuning LLMs on synthetic data that is constructed to contain consistent explanations. The synthetic data is created by prompting LLMs to generate follow-up questions related to initial question-explanation pairs and then answering these questions in a manner consistent with the initial explanation. The method is evaluated on various question-answering datasets and shows a 10.0% relative improvement in explanation consistency on four finetuning datasets and a 4.5% relative improvement on seven out-of-distribution datasets. The paper also discusses the limitations and future directions of EC-finetuning, including its potential for improving explanation consistency in smaller models and its application to more complex tasks.The paper "Towards Consistent Natural-Language Explanations via Explanation-Consistency Finetuning" addresses the issue of inconsistent explanations generated by large language models (LLMs). The authors propose a method called Explanation-Consistency Finetuning (EC-finetuning) to improve the consistency of LLMs' natural-language explanations across related examples. EC-finetuning involves finetuning LLMs on synthetic data that is constructed to contain consistent explanations. The synthetic data is created by prompting LLMs to generate follow-up questions related to initial question-explanation pairs and then answering these questions in a manner consistent with the initial explanation. The method is evaluated on various question-answering datasets and shows a 10.0% relative improvement in explanation consistency on four finetuning datasets and a 4.5% relative improvement on seven out-of-distribution datasets. The paper also discusses the limitations and future directions of EC-finetuning, including its potential for improving explanation consistency in smaller models and its application to more complex tasks.