Crapness Ratio

Crapness Ratio is a metric used to evaluate the proportion of nonsense or invalid information in the answers given by large language models (LLMs). This concept was proposed by Fields Medal winner and Director of Research at Cambridge University Timothy Gowers in 2024. When he tried to use GPT-4o to solve the animal crossing the river problem, he proposed this benchmark, which is the ratio between the total answers given by LLM and the correct answers. Pointing out GPT-4o's mistakes on the simplest questions, Claude 3.5 is not immune. This phenomenon raises questions about whether large language models are really capable of reasoning and planning. Through this ratio, the degree of nonsense output by the model can be quantified, thereby evaluating the performance of the model and the direction of improvement.