Improving the Faithfulness of LLM-based Abstractive Summarization with Span-level Unlikelihood Training

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, medium

Summary

A novel fine-tuning approach enhances the faithfulness of LLM-based abstractive summarization by reducing hallucinations – unfaithful words, phrases, or concepts. Researchers Sicong Huang, Qianqi Yan, Shengze Wang, and Ian Lane developed a method that first annotates span-level hallucinations in LLM-generated summaries using GPT-4 for automatic labeling. Subsequently, the Large Language Model is fine-tuned using both summaries devoid of hallucinations and the identified hallucinated text spans. This technique, termed unlikelihood training, significantly improved summary faithfulness. Experimental results demonstrate a reduction in summaries with hallucinations from 31% to 13% (a 58% decrease) on the CNN summarization dataset, and from 33% to 20% (a 39% decrease) on the SAMSum dataset. The study also introduces a new dataset for distinguishing faithful and unfaithful content and compares unlikelihood training against gradient ascent and task vector negation.

Key takeaway

For NLP Engineers developing abstractive summarization systems, this research offers a direct path to mitigate hallucinations. You should consider implementing span-level unlikelihood training, potentially using large models like GPT-4 for automated annotation of unfaithful text. This approach can substantially reduce hallucination rates, improving summary trustworthiness and user confidence in your LLM-generated content. Explore the provided dataset for distinguishing faithful and unfaithful content to refine your models.

Key insights

Unlikelihood training with span-level hallucination annotations significantly improves LLM abstractive summarization faithfulness.

Principles

• LLM unfaithfulness manifests as span-level hallucinations.
• Fine-tuning can directly target hallucinated text spans.
• Automatic labeling can create training data for faithfulness.

Method

Annotate span-level hallucinations using GPT-4, then fine-tune LLMs with unlikelihood training on both hallucination-free summaries and identified unfaithful spans to reduce their occurrence.

In practice

• Use GPT-4 for automated hallucination annotation.
• Apply unlikelihood training for summarization fine-tuning.
• Develop datasets distinguishing faithful/unfaithful content.

Topics

• LLM Summarization
• Abstractive Summarization
• Hallucination Mitigation
• Unlikelihood Training
• Span-level Annotation
• GPT-4

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Summarisation and Knowledge Distillation — How Agents Summarise Without Hallucinating — Robust LLM summarization requires architectural patterns to prevent hallucinations and ensure factual grounding and completeness.
• Faithful by Construction: Claim-Anchored Attribution for Multi-Document Summarization — A modular, claim-anchored summarization framework inherently builds fine-grained attribution and faithfulness into the generation process.
• LLMs believe false statements even after explicit warnings that they're false — LLMs prioritize statistical patterns in training data over explicit negation, leading to "belief implantation."
• NTS-CoT: Mitigating Hallucinations in LLM-based News Timeline Summarization with Chain-of-Thought Reasoning — Chain-of-Thought reasoning effectively mitigates hallucinations in LLM-based news timeline summarization by structuring information extraction and causal inference.
• MoRFI: Monotonic Sparse Autoencoder Feature Identification — Fine-tuning LLMs on new facts disrupts existing knowledge retrieval, increasing hallucinations via specific latent directions.
• Evidence Graph Consistency in Retrieval-Augmented Generation: A Model-Dependent Analysis of Hallucination Detection — Graph-based RAG hallucination detection is model-dependent, reversing for GPT-class models due to differing hallucination styles.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.