Rephrase and Respond: Let Large Language Models Ask Better Questions for Themselves (RaR) * https://twitter.com/QuanquanGu/status/1722364144379396513 * https://arxiv.org/abs/2311.04205 * https://www.laplace-ai.com/post/thesis-gpt-4-rephrase-and-respond ### One-step RaR prompt: Rephrase and expand the question, and respond. ### Two-step RaR prompt1 : Given the above question, rephrase and expand it to help you do better answering. Maintain all information in the original question. prompt2: (original) {question} (rephrased) {rephrased_question} Use your answer for the rephrased question to answer the original question. ## 有點相關議題 Catch me if you can! How to beat GPT-4 with a 13B model https://lmsys.org/blog/2023-11-14-llm-decontaminator/ 如果 training datasets 裡面已經有 test dataset 資料，那麼 benchmark 會不準過於樂觀啊 不一定是一模一樣或是類似 data，如果透過 rephrase 改寫的話，只用 n-gram overlay 或是相似性比對，就無法檢查出來其實 test dataset 有跟 training dataset 重複喔。 這篇 blog 則是講可用 LLM decontaminator 來檢查