Given the high costs and slow speed of training large language models (LLMs), there is an ongoing discussion about whether spending more compute cycles on inference can help improve the performance of ...