@inproceedings{belem-etal-2025-single,
title = "From Single to Multi: How {LLM}s Hallucinate in Multi-Document Summarization",
author = "Bel{\'e}m, Catarina G and
Pezeshkpour, Pouya and
Iso, Hayate and
Maekawa, Seiji and
Bhutani, Nikita and
Hruschka, Estevam",
editor = "Chiruzzo, Luis and
Ritter, Alan and
Wang, Lu",
booktitle = "Findings of the Association for Computational Linguistics: NAACL 2025",
month = apr,
year = "2025",
address = "Albuquerque, New Mexico",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.findings-naacl.293/",
doi = "10.18653/v1/2025.findings-naacl.293",
pages = "5276--5309",
ISBN = "979-8-89176-195-7",
abstract = "Although many studies have investigated and reduced hallucinations in large language models (LLMs) for single-document tasks, research on hallucination in multi-document summarization (MDS) tasks remains largely unexplored. Specifically, it is unclear how the challenges arising from handling multiple documents (e.g., repetition and diversity of information) affect models outputs. In this work, we investigate how hallucinations manifest in LLMs when summarizing topic-specific information from a set of documents. Since no benchmarks exist for investigating hallucinations in MDS, we leverage existing news and conversation datasets, annotated with topic-specific insights, to create two novel multi-document benchmarks. When evaluating 5 LLMs on our benchmarks, we observe that on average, up to 75{\%} of the content in LLM-generated summary is hallucinated, with hallucinations more likely to occur towards the end of the summaries. Moreover, when summarizing non-existent topic-related information, GPT-3.5-turbo and GPT-4o still generate summaries about 79.45{\%} and 44{\%} of the time, raising concerns about their tendency to fabricate content. To better understand the characteristics of these hallucinations, we conduct a human evaluation of 700+ insights and discover that most errors stem from either failing to follow instructions or producing overly generic insights. Motivated by these observations, we investigate the efficacy of simple post-hoc baselines in mitigating hallucinations but find them only moderately effective. Our results underscore the need for more effective approaches that systematically mitigate hallucinations in MDS."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="belem-etal-2025-single">
<titleInfo>
<title>From Single to Multi: How LLMs Hallucinate in Multi-Document Summarization</title>
</titleInfo>
<name type="personal">
<namePart type="given">Catarina</namePart>
<namePart type="given">G</namePart>
<namePart type="family">Belém</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Pouya</namePart>
<namePart type="family">Pezeshkpour</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hayate</namePart>
<namePart type="family">Iso</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Seiji</namePart>
<namePart type="family">Maekawa</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nikita</namePart>
<namePart type="family">Bhutani</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Estevam</namePart>
<namePart type="family">Hruschka</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2025-04</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Findings of the Association for Computational Linguistics: NAACL 2025</title>
</titleInfo>
<name type="personal">
<namePart type="given">Luis</namePart>
<namePart type="family">Chiruzzo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alan</namePart>
<namePart type="family">Ritter</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lu</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Albuquerque, New Mexico</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-195-7</identifier>
</relatedItem>
<abstract>Although many studies have investigated and reduced hallucinations in large language models (LLMs) for single-document tasks, research on hallucination in multi-document summarization (MDS) tasks remains largely unexplored. Specifically, it is unclear how the challenges arising from handling multiple documents (e.g., repetition and diversity of information) affect models outputs. In this work, we investigate how hallucinations manifest in LLMs when summarizing topic-specific information from a set of documents. Since no benchmarks exist for investigating hallucinations in MDS, we leverage existing news and conversation datasets, annotated with topic-specific insights, to create two novel multi-document benchmarks. When evaluating 5 LLMs on our benchmarks, we observe that on average, up to 75% of the content in LLM-generated summary is hallucinated, with hallucinations more likely to occur towards the end of the summaries. Moreover, when summarizing non-existent topic-related information, GPT-3.5-turbo and GPT-4o still generate summaries about 79.45% and 44% of the time, raising concerns about their tendency to fabricate content. To better understand the characteristics of these hallucinations, we conduct a human evaluation of 700+ insights and discover that most errors stem from either failing to follow instructions or producing overly generic insights. Motivated by these observations, we investigate the efficacy of simple post-hoc baselines in mitigating hallucinations but find them only moderately effective. Our results underscore the need for more effective approaches that systematically mitigate hallucinations in MDS.</abstract>
<identifier type="citekey">belem-etal-2025-single</identifier>
<identifier type="doi">10.18653/v1/2025.findings-naacl.293</identifier>
<location>
<url>https://aclanthology.org/2025.findings-naacl.293/</url>
</location>
<part>
<date>2025-04</date>
<extent unit="page">
<start>5276</start>
<end>5309</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T From Single to Multi: How LLMs Hallucinate in Multi-Document Summarization
%A Belém, Catarina G.
%A Pezeshkpour, Pouya
%A Iso, Hayate
%A Maekawa, Seiji
%A Bhutani, Nikita
%A Hruschka, Estevam
%Y Chiruzzo, Luis
%Y Ritter, Alan
%Y Wang, Lu
%S Findings of the Association for Computational Linguistics: NAACL 2025
%D 2025
%8 April
%I Association for Computational Linguistics
%C Albuquerque, New Mexico
%@ 979-8-89176-195-7
%F belem-etal-2025-single
%X Although many studies have investigated and reduced hallucinations in large language models (LLMs) for single-document tasks, research on hallucination in multi-document summarization (MDS) tasks remains largely unexplored. Specifically, it is unclear how the challenges arising from handling multiple documents (e.g., repetition and diversity of information) affect models outputs. In this work, we investigate how hallucinations manifest in LLMs when summarizing topic-specific information from a set of documents. Since no benchmarks exist for investigating hallucinations in MDS, we leverage existing news and conversation datasets, annotated with topic-specific insights, to create two novel multi-document benchmarks. When evaluating 5 LLMs on our benchmarks, we observe that on average, up to 75% of the content in LLM-generated summary is hallucinated, with hallucinations more likely to occur towards the end of the summaries. Moreover, when summarizing non-existent topic-related information, GPT-3.5-turbo and GPT-4o still generate summaries about 79.45% and 44% of the time, raising concerns about their tendency to fabricate content. To better understand the characteristics of these hallucinations, we conduct a human evaluation of 700+ insights and discover that most errors stem from either failing to follow instructions or producing overly generic insights. Motivated by these observations, we investigate the efficacy of simple post-hoc baselines in mitigating hallucinations but find them only moderately effective. Our results underscore the need for more effective approaches that systematically mitigate hallucinations in MDS.
%R 10.18653/v1/2025.findings-naacl.293
%U https://aclanthology.org/2025.findings-naacl.293/
%U https://doi.org/10.18653/v1/2025.findings-naacl.293
%P 5276-5309
Markdown (Informal)
[From Single to Multi: How LLMs Hallucinate in Multi-Document Summarization](https://aclanthology.org/2025.findings-naacl.293/) (Belém et al., Findings 2025)
ACL