This paper introduces a novel post-hoc refinement method called DETECT, CRITIQUE, and REFINE (DCR) for improving the factual consistency of document-grounded responses generated by large language models (LLMs). The DCR method decomposes the refinement process into three distinct steps: DETECT, CRITIQUE, and REFINE. The DETECT step identifies erroneous generations at the sentence level, the CRITIQUE step generates fine-grained natural language feedback about the errors, and the REFINE step refines the original outputs using the feedback. The authors demonstrate that models of different capabilities benefit from this three-stage refinement approach, outperforming existing end-to-end refinement methods. They evaluate their approach on two datasets: TofuEval and a subset of UltraChat, showing significant improvements in factual consistency. The paper also highlights the importance of the DETECT and CRITIQUE steps in enhancing the post-hoc refinement capabilities of models. The authors fine-tune models to generate fine-grained factual inconsistency localization, reasoning about errors, and suggested fixes, and show that their proposed method produces more effective feedback compared to existing models. The paper concludes by discussing the limitations and potential extensions of the approach.This paper introduces a novel post-hoc refinement method called DETECT, CRITIQUE, and REFINE (DCR) for improving the factual consistency of document-grounded responses generated by large language models (LLMs). The DCR method decomposes the refinement process into three distinct steps: DETECT, CRITIQUE, and REFINE. The DETECT step identifies erroneous generations at the sentence level, the CRITIQUE step generates fine-grained natural language feedback about the errors, and the REFINE step refines the original outputs using the feedback. The authors demonstrate that models of different capabilities benefit from this three-stage refinement approach, outperforming existing end-to-end refinement methods. They evaluate their approach on two datasets: TofuEval and a subset of UltraChat, showing significant improvements in factual consistency. The paper also highlights the importance of the DETECT and CRITIQUE steps in enhancing the post-hoc refinement capabilities of models. The authors fine-tune models to generate fine-grained factual inconsistency localization, reasoning about errors, and suggested fixes, and show that their proposed method produces more effective feedback compared to existing models. The paper concludes by discussing the limitations and potential extensions of the approach.