Chain-of-Thought Emergence › Summary

Chain-of-thought reasoning is one of the most significant emergent capabilities in large language models. By prompting models to generate intermediate reasoning steps, we can access problem-solving abilities that direct questioning leaves unused.

The key insights from this chapter include:

• Scale threshold: CoT benefits emerge only in models above approximately 10 billion parameters. Below this threshold, models produce incoherent reasoning chains that give no benefit.

• Multiple elicitation methods: Few-shot demonstrations, zero-shot triggers like "let's think step by step," and ensemble methods like self-consistency all let CoT reasoning with different tradeoffs.

• Computation extension: The primary mechanism appears to be that intermediate tokens extend the model's effective computation, letting it to solve problems requiring more sequential steps than a single forward pass permits.

• Faithfulness concerns: Generated reasoning chains may not reflect actual internal computation. This limits the reliability of CoT for interpretability and verification.

• Task dependency: CoT helps most on problems with clear decomposition into sequential steps. Not all tasks benefit, and some may be harmed by forced decomposition.

The emergence of chain-of-thought reasoning raises questions about the nature of reasoning in neural networks. Are these models truly reasoning, or merely creating text that mimics reasoning? The next chapter examines how our choice of metrics affects whether capabilities appear emergent, shedding light on this question from a measurement perspective. Later, when we explore inference-time scaling techniques, we'll return to chain-of-thought as a foundation for more advanced reasoning approaches.