Oxford’s AI技術解說 (中英對照)
這份論文叫作《 基礎模型知道如何推理,思考模型只是學會了時機 》(Base Models Know How to Reason, Thinking Models Learn When)。 基礎模型已經擁有推理能力,專家模型只是學會了“何時”使用它。 序章:內力深厚,卻未出招 Prologue: Deep Inner Power, Yet No Moves 在AI的江湖中,基礎模型就像一位閉關多年的武林高手。它在預訓練階段已經習得無數招式——也就是推理的核心知識與模式。 但它的內力雖深,卻未曾真正出招。為什麼?因為它還沒打通任督二脈。 In the realm of AI, a foundation model resembles a martial arts master who has trained in solitude. Through pretraining, it has absorbed countless techniques—core reasoning patterns and knowledge. Yet despite its immense inner power, it doesn’t know when or how to strike. Why? Because its meridians remain blocked. 心法口訣:引導即開關 The Secret Manual: Prompting as the Activation Switch 牛津大學的研究指出,我們無需改變模型的權重(不動筋骨),只要給它一本「心法口訣」——也就是精心設計的Prompt,就能喚醒它沉睡的推理能力。 Oxford’s research reveals that we don’t need to alter the model’s weights (its bones and muscles). Instead, by offering it a “secret manual”—a well-crafted prompt—we can awaken its dormant reasoning power. Chain-of-Thought Prompting :拆招成式,內力外顯 Chain-of-Thought Promp...