Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

Large language models (LLMs) have shown remarkable capabilities in many tasks, but they often struggle with multi-step reasoning problems like math word problems or symbolic manipulation. This paper investigates whether providing models with a chain of thought—a series of intermediate natural language reasoning steps—can unlock these latent reasoning abilities without requiring additional training or fine-tuning.

The authors introduce chain-of-thought prompting, a simple method where a few examples of reasoning are included in the prompt provided to the model. Instead of just giving the model an input-output pair, the prompt includes a triple: 〈input, chain of thought, output〉. The chain of thought mimics a human-like step-by-step thought process, breaking down complex problems into smaller, manageable parts before arriving at the final answer.

The study demonstrates that chain-of-thought prompting leads to significant performance gains across various benchmarks, including arithmetic (GSM8K), commonsense (StrategyQA), and symbolic reasoning (coin flip). A key finding is that this reasoning ability is an emergent property of model scale. While small models often produce fluent but illogical chains of thought, models with approximately 100 billion parameters or more show a dramatic improvement in performance. On the GSM8K benchmark, the PaLM 540B model using chain-of-thought prompting achieved state-of-the-art accuracy, outperforming even models that had been specifically fine-tuned for the task.

This work challenges the notion that scaling up model size alone is sufficient for complex reasoning. By simply changing how we prompt these models, we can elicit sophisticated reasoning behaviors that were previously thought to be out of reach. This approach is highly flexible, as it does not require costly fine-tuning or large labeled datasets, and it provides a window into the model's decision-making process, making its outputs more interpretable.