# An Empirical Comparison of Text Summarization: A Multi-Dimensional Evaluation of Large Language Models

arXiv: https://arxiv.org/abs/2504.04534 · HTML: https://mlsanity.com/paper/2504.04534/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 8 of its 12 references that are in the map.
- First posted on arXiv 6 Apr 2025.

## Builds on (most cited first, 8 of 8)

- [BERTScore: Evaluating Text Generation with BERT](https://mlsanity.com/paper/1904.09675/) (2019)
- [TruthfulQA: Measuring How Models Mimic Human Falsehoods](https://mlsanity.com/paper/2109.07958/) (2021)
- [SummaC: Re-Visiting NLI-based Models for Inconsistency Detection in Summarization](https://mlsanity.com/paper/2111.09525/) (2021)
- [Survey on Factuality in Large Language Models: Knowledge, Retrieval and Domain-Specificity](https://mlsanity.com/paper/2310.07521/) (2023)
- [An Empirical Survey on Long Document Summarization: Datasets, Models and Metrics](https://mlsanity.com/paper/2207.00939/) (2022)
- [Neural Abstractive Text Summarization with Sequence-to-Sequence Models](https://mlsanity.com/paper/1812.02303/) (2018)
- [Factual Consistency Evaluation for Text Summarization via Counterfactual Estimation](https://mlsanity.com/paper/2108.13134/) (2021)
- [A Comparative Study of Quality Evaluation Methods for Text Summarization](https://mlsanity.com/paper/2407.00747/) (2024)

## Abstract (arXiv)

Text summarization is crucial for mitigating information overload across domains like journalism, medicine, and business. This research evaluates summarization performance across 17 large language models (OpenAI, Google, Anthropic, open-source) using a novel multi-dimensional framework. We assessed models on seven diverse datasets (BigPatent, BillSum, CNN/DailyMail, PubMed, SAMSum, WikiHow, XSum) at three output lengths (50, 100, 150 tokens) using metrics for factual consistency, semantic similarity, lexical overlap, and human-like quality, while also considering efficiency factors. Our findings reveal significant performance differences, with specific models excelling in factual accuracy (deepseek-v3), human-like quality (claude-3-5-sonnet), and processing efficiency/cost-effectiveness (gemini-1.5-flash, gemini-2.0-flash). Performance varies dramatically by dataset, with models struggling on technical domains but performing well on conversational content. We identified a critical tension between factual consistency (best at 50 tokens) and perceived quality (best at 150 tokens). Our analysis provides evidence-based recommendations for different use cases, from high-stakes applications requiring factual accuracy to resource-constrained environments needing efficient processing. This comprehensive approach enhances evaluation methodology by integrating quality metrics with operational considerations, incorporating trade-offs between accuracy, efficiency, and cost-effectiveness to guide model selection for specific applications.