#推理机制

Understanding these three mechanisms helps demystify how large language models work and why they can answer questions without explicit memorization.

Google用PaLM 2发现，模型不靠堆算力背答案，而是靠推理串起脑袋里分散的知识点。比直接猜准多了。