The “reasoning” which is displayed, such as step-by-step explanations or chain-of-thought responses, is only a simulation of reasoning, generated as text based on patterns. It’s in the same form of any text production an LLM is capable of. When an LLM mimics “reasoning,” it is generating text that looks like reasoning because it has seen similar patterns in its training data.
Anthropic have actually looked at how their LLMs reason. Don’t use them for anything important.
The “reasoning” which is displayed, such as step-by-step explanations or chain-of-thought responses, is only a simulation of reasoning, generated as text based on patterns. It’s in the same form of any text production an LLM is capable of. When an LLM mimics “reasoning,” it is generating text that looks like reasoning because it has seen similar patterns in its training data.