Enhancing Faithfulness in Abstractive Summarization via Span-Level Fine-Tuning

2025-10-10

AI Summary

To address pervasive hallucination in abstractive summarization by large language models (LLMs), this paper proposes a fine-tuning framework grounded in segment-level faithfulness annotation. First, GPT-4o is employed to perform fine-grained annotation—labeling hallucinations at the token, phrase, and conceptual levels—yielding the first summarization dataset with explicit, hierarchical hallucination annotations. Second, three novel supervised fine-tuning strategies are introduced: non-likelihood training, gradient ascent, and task vector negation. Experiments demonstrate that all three methods significantly improve summary faithfulness; non-likelihood training achieves the strongest performance, outperforming baselines across all granularity levels. This work constitutes the first effort to model and optimize for hallucination at the segment level, offering both a scalable methodology and a high-quality benchmark resource to enhance the factual consistency of LLM-generated summaries.

Technology Category

• Natural Language Processing: Summarization
• Machine Learning: Large Multimodal Models (LMMs)
• Data Mining & Knowledge Management: Data Visualization & Summarization

Application Category

• Search and Retrieval-Augmented AI: Large language models for search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling

Abstract

Abstractive summarization using large language models (LLMs) has become an essential tool for condensing information. However, despite their ability to generate fluent summaries, these models sometimes produce unfaithful summaries, introducing hallucinations at the word, phrase, or concept level. Existing mitigation strategies, such as post-processing corrections or contrastive learning with synthetically generated negative samples, fail to fully address the diverse errors that can occur in LLM-generated summaries. In this paper, we investigate fine-tuning strategies to reduce the occurrence of unfaithful spans in generated summaries. First, we automatically generate summaries for the set of source documents in the training set with a variety of LLMs and then use GPT-4o to annotate any hallucinations it detects at the span-level. Leveraging these annotations, we fine-tune LLMs with both hallucination-free summaries and annotated unfaithful spans to enhance model faithfulness. In this paper, we introduce a new dataset that contains both faithful and unfaithful summaries with span-level labels and we evaluate three techniques to fine-tuning a LLM to improve the faithfulness of the resulting summarization: gradient ascent, unlikelihood training, and task vector negation. Experimental results show that all three approaches successfully leverage span-level annotations to improve faithfulness, with unlikelihood training being the most effective.

Problem

• Reducing hallucinations in abstractive summarization models
• Addressing unfaithful spans in LLM-generated summaries
• Improving faithfulness through span-level fine-tuning techniques

Innovation

• Fine-tuning LLMs with span-level hallucination annotations
• Using GPT-4o to automatically label unfaithful text spans
• Applying unlikelihood training to reduce summary hallucinations