Researchers have investigated the effect of pruning on large language models (LLMs) used for reasoning tasks. They found that unstructured pruning, which removes individual weights rather than entire layers, can actually improve performance compared to structured pruning and even outperform unpruned models in some cases. This challenges the conventional understanding that pruning always reduces test-time scaling effectiveness. --- Why it matters: This matters because it suggests that carefully applied pruning techniques could be used to make LLMs more efficient without sacrificing their reasoning capabilities, which is important for applications like natural language processing and question-answering systems. Source: https://arxiv.org/abs/2604.25098