Gemini surpasses GPT-4 on MMLU tasks with carefully crafted prompts

The higher the better

CoT@32: a method where LLMs uses 32 reasoning steps, picks the answer with the highest agreement among them, or defaults to a single step if there's no clear consensus. 5-shot indicates the number of examples given to the LLM for each specific task during the evaluation. Think of it as giving a student five practice problems before administering the actual test.
Source: DeepMind paperCreated with Datawrapper