XL$^2$Bench: A Benchmark for Extremely Long Context Understanding with Long-range Dependencies
This addresses the need for better evaluation tools for long-context LLMs, which is crucial for researchers and developers working on applications requiring extensive text comprehension, though it is incremental as it builds on prior benchmarking efforts.
The paper introduces XL$^2$Bench, a benchmark for evaluating large language models on extremely long context understanding with long-range dependencies, covering scenarios like fiction, paper, and law reading with tasks up to 100K+ words, and finds that six leading LLMs perform significantly worse than humans.
Large Language Models (LLMs) have demonstrated remarkable performance across diverse tasks but are constrained by their small context window sizes. Various efforts have been proposed to expand the context window to accommodate even up to 200K input tokens. Meanwhile, building high-quality benchmarks with much longer text lengths and more demanding tasks to provide comprehensive evaluations is of immense practical interest to facilitate long context understanding research of LLMs. However, prior benchmarks create datasets that ostensibly cater to long-text comprehension by expanding the input of traditional tasks, which falls short to exhibit the unique characteristics of long-text understanding, including long dependency tasks and longer text length compatible with modern LLMs' context window size. In this paper, we introduce a benchmark for extremely long context understanding with long-range dependencies, XL$^2$Bench, which includes three scenarios: Fiction Reading, Paper Reading, and Law Reading, and four tasks of increasing complexity: Memory Retrieval, Detailed Understanding, Overall Understanding, and Open-ended Generation, covering 27 subtasks in English and Chinese. It has an average length of 100K+ words (English) and 200K+ characters (Chinese). Evaluating six leading LLMs on XL$^2$Bench, we find that their performance significantly lags behind human levels. Moreover, the observed decline in performance across both the original and enhanced datasets underscores the efficacy of our approach to mitigating data contamination.