why does performance improve after chain of thought prompting? Because a human is measuring it unfairly. The output without CoT is valid. It is syntactically

why does performance improve after chain of thought prompting? Because a human is measuring it unfairly. The output without CoT is valid. It is syntactically valid. The observer is unhappy with the semantic validity, because the observer has seen syntactic validity and assumed that semantic validity is a given. Like it would if the model was alive. This is observer error, not model error.