commonsense-benchmarks

This paper tests whether commonsense benchmark scores predict real-world downstream task performance by evaluating 23 LLMs across four benchmarks and their reworked variants, finding that revisions preserve rankings but only offer task-dependent predictive validity.