Chain-of-Thought Emergence › The Discovery of Chain-of-Thought

Chain-of-thought prompting was systematically studied by Wei et al. (2022), though researchers had observed similar phenomena informally. The core finding was deceptively simple yet highly significant: including step-by-step reasoning examples in the prompt causes models to generate their own intermediate reasoning steps, sharply improving performance on tasks requiring multi-step inference. This discovery revealed that the barrier to complex reasoning wasn't necessarily the model's lack of capability, but rather our failure to communicate effectively with these systems about how we wanted them to approach problems.

To understand why this matters, consider how humans naturally solve complex problems. We don't typically leap directly from question to answer. Instead, we break problems into manageable pieces, work through each piece systematically, and combine our intermediate results. A student solving a word problem writes out their calculations. A detective articulates their reasoning before reaching a conclusion. Chain-of-thought prompting in effect teaches language models to adopt this same structured approach.

Consider a standard arithmetic word problem that illustrates the challenge:

Roger has 5 tennis balls. He buys 2 more cans of tennis balls. Each can has 3 balls. How many tennis balls does he have now?

A model prompted with just the question might output "11" directly, with no explanation. Often this answer would be wrong. The model might say "10" by confusing the number of cans with the total balls purchased, or "8" by only partially accounting for the original balls. The problem isn't that the model lacks knowledge of arithmetic or the meaning of the words. Rather, the model struggles to orchestrate its knowledge into the correct sequence of operations when asked to produce an immediate answer.

But when shown examples where each step is made explicit, something remarkable happens:

Roger has 5 tennis balls. He buys 2 more cans of tennis balls. Each can has 3 balls. How many tennis balls does he have now?

Roger started with 5 balls. 2 cans at 3 balls each is 6 balls. 5 + 6 = 11. The answer is 11.

The model learns to generate similar reasoning chains for new problems, and its accuracy improves substantially. By seeing that the appropriate response format involves articulating intermediate calculations, the model begins to produce its own step-by-step reasoning when faced with novel questions.

What makes this surprising is that no one explicitly trained these models to reason step-by-step. Standard language model pretraining involves predicting the next token in text from the internet and other corpora. The training objective is simply to become better at predicting what word comes next. There's no explicit reward signal for logical reasoning or mathematical accuracy. Yet somewhere in this process, models acquire the ability to decompose problems and work through them sequentially, a capability that only manifests when properly prompted. The reasoning ability appears to be learned implicitly, as a byproduct of exposure to human text containing worked explanations and solutions.

The original observations about CoT emergence came from several key findings that together painted a picture of a novel capability. First, smaller models showed no improvement (or even degradation) when prompted with chain-of-thought examples. Models with fewer than approximately 10 billion parameters often produced incoherent reasoning chains or simply ignored the demonstration format entirely. When smaller models attempted to generate reasoning steps, the results were frequently nonsensical: steps that did not follow logically from each other, arithmetic operations applied incorrectly, or conclusions that bore no relationship to the preceding "reasoning." This was not just poor performance. It was qualitatively different from what larger models produced.

Second, large models showed large improvements that exceeded what anyone had anticipated. PaLM 540B achieved 57.1% accuracy on the GSM8K math benchmark with CoT prompting compared to just 17.9% with with standard prompting, a more than threefold improvement from the same model using the same knowledge, differing only in how the question was posed. This improvement wasn't a modest gain that might be attributed to noise or lucky prompt engineering; it was a major change in the model's apparent capabilities.

Third, the reasoning chains themselves were often (though not always) logically coherent and human-interpretable. Models were not just creating word salad that happened to precede correct answers; they were generating valid intermediate steps that a human could follow and verify. Each step typically followed from the previous one, arithmetic operations were usually applied correctly, and the logical structure of the argument was sound. This interpretability meant that when models made errors, researchers could often identify exactly where the reasoning went wrong. This transparency is something direct answers cannot give.

The phase transition aspect was striking and connects to important questions about how capabilities develop in neural networks. As we discussed in the chapter on emergence in neural networks, emergent capabilities often appear suddenly at certain scale thresholds rather than improving gradually. CoT reasoning follows this pattern with remarkable clarity: models below a necessary size gain nothing from chain-of-thought prompting, while models above this threshold show substantial benefits. The transition isn't gradual. There's a relatively narrow band of model sizes where CoT effectiveness goes from in effect zero to sharply positive.