📄 中文摘要
2026年2月17日，xAI发布了Grok 4.20的公开测试版，其主要特点是四个专门的AI代理——船长、研究员、逻辑学家和创意者，它们在产生最终答案之前进行辩论。这一方法使得幻觉率从12%降至4.2%。然而，Grok的方法未能解决一个更深层次的问题：如果这种辩论发生在单一模型内部，而不是在不同模型之间，会怎样？为此，MDIR（多深度迭代推理）架构应运而生，它在单一的变换器骨干中嵌入多个认知处理器。该架构旨在提高推理的深度和准确性。
MDIR：如果一个模型能够在回答之前自我辩论会怎样？
出处: MDIR: What If a Single Model Could Debate Itself Before Answering?
发布: 2026年2月22日
📄 English Summary
MDIR: What If a Single Model Could Debate Itself Before Answering?
On February 17, 2026, xAI released the public beta of Grok 4.20, featuring four specialized AI agents—a Captain, a Researcher, a Logician, and a Creative—that debate each other before producing a final answer. This approach reduced hallucination rates from 12% to 4.2%. However, it raises a deeper question: what if this deliberation occurred within a single model instead of between separate ones? This is the premise behind MDIR (Multi-Depth Iterative Reasoning), an architecture designed to embed multiple cognitive processors within a single transformer backbone, aiming to enhance the depth and accuracy of reasoning.