LongBench v2

503 hard multiple-choice questions with contexts from 8k to 2M words, built so a model must reason over long context rather than just retrieve, with human experts scoring only 53.7%.

Also known as: LongBench v2: Towards Deeper Understanding and Reasoning on Realistic Long-context Multitasks

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.
Categorylong-context
Subcategorymultiple-choice long-context deep understanding and reasoning (8k-2M words)
Page statusactive
Metricaccuracy
Directionhigher_is_better
Unit%
Dataset size503
Dataset licenceApache-2.0 (per the Hugging Face dataset card)
PublisherTsinghua University and Zhipu.AI

What it measures

LongBench v2 gives a model a long context (8,000 to 2,000,000 words, mostly under 128K) and a four-option multiple-choice question that cannot be answered by simple retrieval. The 503 questions span six task categories: single-document QA, multi-document QA, long in-context learning, long-dialogue history understanding, code repository understanding, and long structured-data (table/graph) understanding. Questions were collected from nearly 100 highly educated contributors across diverse professional backgrounds, then filtered through automated and manual review for quality and genuine difficulty. The explicit design goal is to require deep understanding and multi-step reasoning over the full context rather than locating one planted fact, in contrast to needle-in-a-haystack-style long-context tests.

Task format

A long document, document set, dialogue history, code repository, or structured dataset plus a four-option multiple-choice question in; a single selected option (A-D) out. All questions carry a difficulty label (easy/hard) and a length bucket (short/medium/long), and the reference protocol reports both direct answering and answering with extended chain-of-thought reasoning.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub