CoT moderately improves model performance, but all non-o1 models suck at actually thinking step by step effectively. If the task is not straightforward, they ma

CoT moderately improves model performance, but all non-o1 models suck at actually thinking step by step effectively. If the task is not straightforward, they make obvious mistakes or default to guessing. OpenAI trained o1 to pick better steps in its chain of thought. (The moat is the dataset they used to do that.)