All terms
Evaluation
LongBench
A benchmark suite measuring how well models handle long-context tasks.
Definition
LongBench is a benchmark suite for testing how well language models use long inputs, covering tasks such as multi-document question answering, summarization, and retrieval over passages that span many thousands of tokens. It probes whether a model genuinely reads across its full context window rather than attending only to the start or end. It is used to compare long-context performance across models.