LiveReasonBench

An OpenCompass-maintained, free-response QA benchmark graded like OpenAI's SimpleQA and refreshed through periodic dated dataset versions; its underlying question set is not publicly downloadable.

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.
Categoryreasoning
Subcategoryfree-response general QA/reasoning, periodically re-versioned, SimpleQA-style grading
Page statusactive
Metricaccuracy_given_attempted (precision on attempted answers) and F1, following SimpleQA's own metric design
Directionhigher_is_better
Unit%
PublisherShanghai Artificial Intelligence Laboratory (OpenCompass project)

What it measures

LiveReasonBench poses a single free-response question and compares the model's answer against a gold target, with no answer options offered. Its harness code is adapted directly from OpenAI's SimpleQA grading rubric, down to reusing SimpleQA's own worked examples (the Barack Obama's children example, the Jason Wei height example) inside its LLM-judge prompt. What exact topic or difficulty range the questions cover could not be independently confirmed for this page: the dataset (nominally hosted at `opencompass/LiveReasonBench` on Hugging Face) returns an authentication error to an anonymous request rather than a public dataset card, so its content is not publicly inspectable even though the evaluation code that runs against it is. The name and the question/gold-target/free-response structure suggest general reasoning or knowledge QA rather than a narrow task, but this is inferred from the harness mechanism, not confirmed from example items.

Task format

Single-turn free-response question answering: the model receives only "Question: {question}" with no options, and its answer is graded by a separate LLM-judge call comparing it against a gold target, following the CORRECT / INCORRECT / NOT_ATTEMPTED rubric from OpenAI's SimpleQA.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub