Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Investigating Hallucinations in Pruned Large Language Models for Abstractive Summarization

George Chrysostomou, Zhixue Zhao, Miles Williams, Nikolaos Aletras

TL;DR

Surprisingly, it is found that hallucinations are less prevalent from pruned LLMs than the original models, and analysis suggests that pruned models tend to depend more on the source document for summary generation, which could be a reason for the reduction in hallucination risk.

Abstract

Despite the remarkable performance of generative large language models (LLMs) on abstractive summarization, they face two significant challenges: their considerable size and tendency to hallucinate. Hallucinations are concerning because they erode reliability and raise safety issues. Pruning is a technique that reduces model size by removing redundant weights, enabling more efficient sparse inference. Pruned models yield downstream task performance comparable to the original, making them ideal alternatives when operating on a limited budget. However, the effect that pruning has upon hallucinations in abstractive summarization with LLMs has yet to be explored. In this paper, we provide an extensive empirical study across five summarization datasets, two state-of-the-art pruning methods, and five instruction-tuned LLMs. Surprisingly, we find that hallucinations are less prevalent from pruned LLMs than the original models. Our analysis suggests that pruned models tend to depend more on the source document for summary generation. This leads to a higher lexical overlap between the generated summary and the source document, which could be a reason for the reduction in hallucination risk.

Investigating Hallucinations in Pruned Large Language Models for Abstractive Summarization

TL;DR

Surprisingly, it is found that hallucinations are less prevalent from pruned LLMs than the original models, and analysis suggests that pruned models tend to depend more on the source document for summary generation, which could be a reason for the reduction in hallucination risk.

Abstract

Despite the remarkable performance of generative large language models (LLMs) on abstractive summarization, they face two significant challenges: their considerable size and tendency to hallucinate. Hallucinations are concerning because they erode reliability and raise safety issues. Pruning is a technique that reduces model size by removing redundant weights, enabling more efficient sparse inference. Pruned models yield downstream task performance comparable to the original, making them ideal alternatives when operating on a limited budget. However, the effect that pruning has upon hallucinations in abstractive summarization with LLMs has yet to be explored. In this paper, we provide an extensive empirical study across five summarization datasets, two state-of-the-art pruning methods, and five instruction-tuned LLMs. Surprisingly, we find that hallucinations are less prevalent from pruned LLMs than the original models. Our analysis suggests that pruned models tend to depend more on the source document for summary generation. This leads to a higher lexical overlap between the generated summary and the source document, which could be a reason for the reduction in hallucination risk.
Paper Structure (41 sections, 6 equations, 2 figures, 8 tables)

This paper contains 41 sections, 6 equations, 2 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Hallucinations in Summarization
  4. Measuring Hallucination Risk
  5. Pruning Large Language Models
  6. Methodology
  7. Models
  8. Pruning Methods
  9. Magnitude hagiwara-1994-simplehan-etal-2015-learning
  10. SparseGPT frantar-alistarh-2023-sparsegpt
  11. Wanda sun-etal-2024-simple
  12. Sparsity Level
  13. Prompting
  14. Summarization Datasets
  15. Evaluation of Summarization Quality
  16. ...and 26 more sections

Figures (2)

  • Figure 1: An example of a hallucination (highlighted text) in abstractive summarization.
  • Figure 2: Ratio between a pruned model and the original across five sparsity levels, three hallucination risk metrics (lines with circled markers; lower means pruned is better) and four summary generation performance metrics (gray dotted lines; higher means pruned is better). The ratio for each metric is averaged across all datasets, with error bars indicating standard deviation.