This paper proposes a three-stage refinement framework for improving factual consistency in document-grounded summaries. The framework consists of three distinct LLM competencies: DETECT, CRITIQUE, and REFINE. The DETECT step identifies erroneous generations at the sentence level. The CRITIQUE step generates fine-grained natural language feedback about errors and how to fix them. The REFINE step uses this feedback to make targeted edits to the original response. The approach is evaluated on two datasets: TofuEval and a subset of UltraChat. The results show that the proposed method consistently outperforms existing end-to-end refinement approaches and current trained models not fine-tuned for factuality critiquing. The method is effective across a range of LLM capabilities, from LLAMA-2-7B-CHAT to GPT-4. The approach uses a detector to focus feedback and allows for fine-tuning models on fine-grained feedback to enumerate specific errors. The method is shown to be more effective than prior approaches that either use general instructions or require verification in the feedback step. The paper also shows that fine-tuning the critique model improves its capabilities over prompting, and the model is able to give feedback on a variety of factual inconsistencies. The main contributions are: (1) introducing a novel post-hoc refinement method: DETECT, CRITIQUE and REFINE (DCR), which refines with natural language feedback to enhance factual consistency; (2) fine-tuning models to generate fine-grained factual inconsistency localization, reasoning about the error, and a suggested fix for the inconsistency; (3) showing the importance of the DETECT and CRITIQUE steps in enhancing the post-hoc refinement capabilities of models. The paper also evaluates the effectiveness of the proposed method against existing feedback models and shows that the proposed feedback leads to higher factual consistency post-refinement than feedback from Shepherd or SelFee. The results show that the three-stage approach outperforms ablations removing or simplifying these stages. The paper also shows that the proposed method is effective across a range of LLM capabilities and that the form of feedback given by the models leads to higher factual consistency post-refinement than feedback from existing models. The paper also discusses the limitations of the approach, including the reliance on an off-the-shelf and reliable DETECT model for sentence-level factual consistency detection and the need for further exploration on how to effectively choose and train M_detect for tasks other than document-grounded factuality detection.This paper proposes a three-stage refinement framework for improving factual consistency in document-grounded summaries. The framework consists of three distinct LLM competencies: DETECT, CRITIQUE, and REFINE. The DETECT step identifies erroneous generations at the sentence level. The CRITIQUE step generates fine-grained natural language feedback about errors and how to fix them. The REFINE step uses this feedback to make targeted edits to the original response. The approach is evaluated on two datasets: TofuEval and a subset of UltraChat. The results show that the proposed method consistently outperforms existing end-to-end refinement approaches and current trained models not fine-tuned for factuality critiquing. The method is effective across a range of LLM capabilities, from LLAMA-2-7B-CHAT to GPT-4. The approach uses a detector to focus feedback and allows for fine-tuning models on fine-grained feedback to enumerate specific errors. The method is shown to be more effective than prior approaches that either use general instructions or require verification in the feedback step. The paper also shows that fine-tuning the critique model improves its capabilities over prompting, and the model is able to give feedback on a variety of factual inconsistencies. The main contributions are: (1) introducing a novel post-hoc refinement method: DETECT, CRITIQUE and REFINE (DCR), which refines with natural language feedback to enhance factual consistency; (2) fine-tuning models to generate fine-grained factual inconsistency localization, reasoning about the error, and a suggested fix for the inconsistency; (3) showing the importance of the DETECT and CRITIQUE steps in enhancing the post-hoc refinement capabilities of models. The paper also evaluates the effectiveness of the proposed method against existing feedback models and shows that the proposed feedback leads to higher factual consistency post-refinement than feedback from Shepherd or SelFee. The results show that the three-stage approach outperforms ablations removing or simplifying these stages. The paper also shows that the proposed method is effective across a range of LLM capabilities and that the form of feedback given by the models leads to higher factual consistency post-refinement than feedback from existing models. The paper also discusses the limitations of the approach, including the reliance on an off-the-shelf and reliable DETECT model for sentence-level factual consistency detection and the need for further exploration on how to effectively choose and train M_detect for tasks other than document-grounded factuality detection.