HelloBench: Evaluating Long Text Generation Capabilities of Large Language Models

TL;DR

HelloBench is a comprehensive benchmark designed to evaluate large language models' ability to generate long texts across various tasks, addressing a gap in current assessments and proposing an efficient human-aligned evaluation method.

Contribution

The paper introduces HelloBench and HelloEval, new benchmarks and evaluation methods specifically for assessing long text generation in LLMs, with extensive experiments on 30 models.

Findings

Most LLMs cannot generate texts longer than 4000 words.

Longer text generation often suffers from repetition and quality issues.

HelloEval correlates highly with human judgment, outperforming traditional metrics.

Abstract

In recent years, Large Language Models (LLMs) have demonstrated remarkable capabilities in various tasks (e.g., long-context understanding), and many benchmarks have been proposed. However, we observe that long text generation capabilities are not well investigated. Therefore, we introduce the Hierarchical Long Text Generation Benchmark (HelloBench), a comprehensive, in-the-wild, and open-ended benchmark to evaluate LLMs' performance in generating long text. Based on Bloom's Taxonomy, HelloBench categorizes long text generation tasks into five subtasks: open-ended QA, summarization, chat, text completion, and heuristic text generation. Besides, we propose Hierarchical Long Text Evaluation (HelloEval), a human-aligned evaluation method that significantly reduces the time and effort required for human evaluation while maintaining a high correlation with human evaluation. We have conducted extensive experiments across around 30 mainstream LLMs and observed that the current LLMs lack long text generation capabilities. Specifically, first, regardless of whether the instructions include explicit or implicit length constraints, we observe that most LLMs cannot generate text that is longer than 4000 words. Second, we observe that while some LLMs can generate longer text, many issues exist (e.g., severe repetition and quality degradation). Third, to demonstrate the effectiveness of HelloEval, we compare HelloEval with traditional metrics (e.g., ROUGE, BLEU, etc.) and LLM-as-a-Judge methods, which show that HelloEval has the highest correlation with human evaluation. We release our code in https://github.com/Quehry/HelloBench.