Tag
This paper reveals a 'benchmark illusion' where pruned LLMs perform well on multiple-choice benchmarks but fail to answer the same questions in open generation, suggesting that compressed models should be tested on generative tasks rather than just recognition tasks.