1
The Benchmark Illusion: Pruned LLMs Can Pass Multiple Choice but Fail to Answer
剪枝后的大模型在多项选择基准测试中表现优异，却在开放回答任务中“原形毕露”，揭示评测陷阱。
arXiv:2606.17609v1 Announce Type: new Abstract: Compressing large language models reduces memory use and inference cost, but it can also create failur…