Learning from Self Critique and Refinement for Faithful LLM Summarization

TL;DR

This paper tackles hallucinations in LLM-based abstractive summarization by introducing SCRPO, a self-supervised framework that leverages the model's own critique and refinement to construct a self-generated preference dataset and train via preference learning. The method eliminates additional test-time computation and demonstrates superior faithfulness on XSum, CNNDM, and SAMSum, while maintaining or improving overall summary quality and showing cross-domain robustness. Key contributions include a dual-strategy critique (binary and fine-grained), a DPO-based training objective with NLL regularization, and evidence that SCRPO outperforms inference-time refinement and other self-supervised baselines. The work has practical implications for deploying faithful summarization systems without external supervision or increased inference cost.

Abstract

Large Language Models (LLMs) often suffer from hallucinations: output content that is not grounded in the input context, when performing long-form text generation tasks such as summarization. Prior works have shown that hallucinations can be reduced by iteratively critiquing and refining previously generated outputs using either the same model or a more powerful teacher model as the critique. However, these approaches either require additional test-time compute or assume access to more powerful teacher models, making them costly and less practical. In this work, we propose Self Critique and Refinement-based Preference Optimization (SCRPO), which is a self-supervised training framework that first constructs a preference dataset by leveraging the LLM's own critique and refinement capabilities, and then applies preference learning to improve the same LLM for faithful summarization. Experiments on three summarization benchmarks (XSUM CNNDM and SAMSum), demonstrate that our approach outperforms state-of-the-art self-supervised learning methods in terms of faithfulness metrics while either maintaining or improving other metrics that measure the overall quality of the summary. Moreover, compared to test-time refinement, our approach not only improves efficiency but also results in more faithful summaries.

Figures (3)

• Figure 1: Overview of the proposed SCRPO framework. Given a set of unlabeled documents $\{x\}$, we use a pretrained LLM $\pi$ to construct a preference dataset of summaries, and then finetune the pretrained LLM with preference learning to improve the faithfulness of generated summaries. The details of preference data construction are elaborated in Algorithm \ref{['alg:ALG1']}
• Figure 2: LLM critique with fine-grained feedback. We prompt the same LLM $\pi$ to perform atomic fact extraction (with prompt $p_{atomic\_fact}$) and natural language inference (with prompt $p_{nli}$). The hallucination score and critique feedback are obtained using the atomic facts that are not entailed.
• Figure 3: Impact of SCRPO training on models of different sizes (dataset - XSum).