CHARM (Benchmarking Chinese Commonsense Reasoning of LLMs)

Tests Chinese-specific commonsense reasoning against a matched globally-known-commonsense control, plus free-form memorization questions built from the same underlying facts.

Also known as: CHARM

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
SubcategoryChinese commonsense reasoning and knowledge memorization
Page statusactive
Metricaccuracy (multiple-choice for reasoning tasks; LLM-judged or rule-matched for memorization tasks)
Directionhigher_is_better
Unit%
Dataset size2559
Dataset licenceApache-2.0
PublisherShanghai AI Laboratory; Tongji University; Wuhan University

What it measures

CHARM evaluates a model's commonsense reasoning in Chinese across two matched domains: globally known commonsense and Chinese-specific commonsense, the latter spanning seven aspects (history, traditional culture and arts, daily life and customs, entertainment, public figures, geography, and the Chinese language itself). The same seven task types -- such as judging historical anachronisms, understanding a time expression, or recommending a movie -- are built in both domains with identical formats, so a model's gap between the two domains isolates how much of its performance depends on Chinese-specific knowledge rather than general reasoning ability. A separate set of memorization tasks, built from the same underlying facts as four of the reasoning tasks, tests whether a model actually knows the relevant facts, to distinguish reasoning failures from simple forgetting.

Task format

14 multiple-choice reasoning subtasks (7 task types, each run in the Chinese-specific and global domains), with 2 to 6 options depending on the subtask, plus 4 free-form question-answering memorization subtasks derived from 4 of the 7 reasoning task types. The original paper evaluated 19 LLMs under 5 prompt strategies, including chain-of-thought and a cross-lingual-thought (XLT) strategy; the public leaderboard reports each model under whichever strategy (XLT for English-oriented models, Chinese chain-of-thought for Chinese-oriented models) scored best for it.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub