The user wants me to rewrite a headline about Chain of Thought prompting. Let me analyze the original:

Short summary

Chain-of-thought prompting works because transformers have a fixed compute budget per token—adding reasoning tokens gives the model multiple forward passes and an external scratchpad to store intermediate results. It helps when problems have serial dependencies (each step needs the prior step's output) but does nothing for single-lookup or one-pass classification tasks. The written reasoning is not an audit log; the model may rationalize a plausible chain after already landing on an answer.

• •Reasoning tokens buy additional forward passes and external memory, not extra intelligence
• •CoT helps for multi-step problems with serial dependencies but is useless for single-pass tasks
• •The written chain may be a post-hoc rationalization, not a replay of actual computation

Generated with AI, which can make mistakes.