OOLONG is a long-context benchmark, released in 2025, that measures how well a language model reads and aggregates information scattered across a long context to answer questions. Unlike needle-in-a-haystack tasks that look for a single passage, its tasks require processing many items in the context one by one and aggregating or reasoning over them.
It is used to show how model scores drop sharply as context length grows and to evaluate long-context processing techniques.