Rethinking Benchmark and Contamination for Language Models with Rephrased Samples

This work studies benchmark contamination under rephrasing and motivates better dataset hygiene and evaluation practices for language models.