咖啡郎
10日
係咪免費版



attern Recognition: LLMs have read vast amounts of math-related text, textbooks, and code during training. They recognize that certain inputs are often followed by certain outputs (e.g., they know "2 + 2" usually goes with "4" simply because they’ve seen it thousands of times).Chain-of-Thought Reasoning: Instead of guessing the final answer directly, LLMs are trained to "show their work". By breaking complex math into a sequence of smaller steps, the model predicts one token at a time, mimicking how a human does long division or algebra on paper.Tool Integration: Modern LLMs act as controllers. When asked to perform a complex calculation (like \(x^2 - 3x + 2 = 0\)), many models will write background programming code (such as Python) and use a calculator/tool to get the exact answer, rather than guessing