Chain of thought" (CoT) prompting in improving the reasoning abilities of LLM

This research explores the effectiveness of "chain of thought" (CoT) prompting in improving the reasoning abilities of large language models (LLMs). The authors posit that CoT empowers LLMs to perform serial computations, a capability that is otherwise lacking in traditional transformer architectures. They analyze the expressiveness of transformers with and without CoT using circuit complexity and establish that constant-depth, constant-precision transformers with CoT can solve problems in the class P/poly, a superclass of P. This finding implies that with polynomially many CoT steps, these transformers can handle problems previously considered intractable for parallel computation. The theoretical analysis is supported by empirical evaluations on tasks that are inherently serial, demonstrating a significant improvement in accuracy when CoT is employed, especially for low-depth transformers.