Chain Of Thoughts Prompting
This research paper explores chain-of-thought prompting, a technique that significantly improves the complex reasoning abilities of large language models (LLMs). By providing LLMs with a few examples of problems solved using a step-by-step reasoning process (chain of thought), the researchers demonstrate substantial performance gains across various reasoning tasks, including arithmetic, commonsense, and symbolic reasoning. The study finds that this improvement is strongly linked to the scale of the LLM, with smaller models showing little to no benefit. The effectiveness of chain-of-thought prompting is also robust across different datasets and annotators, highlighting its potential as a broadly applicable method for enhancing LLM reasoning capabilities. The authors acknowledge limitations concerning the factuality of generated reasoning steps and the cost associated with using very large models.
