Public demo

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

Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, Denny Zhou (2022)

Chain-of-thought prompting shows that large language models can improve on multi-step reasoning tasks when prompted to produce intermediate reasoning steps.

Why it matters

The paper helped turn prompting into a method for eliciting reasoning behavior, not just task instructions. It also exposed a core theme in modern LLM systems: performance can depend strongly on how evidence, steps, and intermediate structure are represented.

Highlights

  • [method] The prompt includes examples with intermediate reasoning chains before the final answer. — why: this simple format change is the main intervention.
  • [finding] The approach improves performance on arithmetic, symbolic, and commonsense reasoning tasks for sufficiently large models. — why: the size dependence is central to the paper's claim.
  • [limitation] Chain-of-thought is less effective for smaller models and does not guarantee faithful reasoning. — why: generated reasoning can still be plausible-looking rather than causally correct.
  • [gap] The paper motivates later work on verification, self-consistency, and tool use. — why: reasoning traces create something to inspect, sample, or check.

Reader notes

  • In a SAV demo, this paper connects directly to evidence-supported writing: intermediate steps are useful only if they can be checked against sources.
  • Compare with toolformer: CoT elicits internal reasoning, while Toolformer lets models call external tools.
  • Good citation target when explaining why "show the steps" is helpful but not enough for trust.

Extracted evidence examples

  • OBJECTIVE: Combining rationale generation with few-shot prompting, this work enhances LLM reasoning by eliciting emergent, sequential thought processes, similar to Chain-of-Thought.
    This work explores how the reasoning ability of large language models can be unlocked by a simple method motivated by two ideas.
    In this paper, we combine the strengths of these two ideas in a way that avoids their limitations.
  • OBJECTIVE: Errors are categorized based on the changes needed to correct the chain of thought, aiming for future model improvement.
    We decided to categorize errors into what changes are needed to make the chain of thought correct, with the goal of elucidating how the model can be improved in the future.
  • RESEARCH_PROBLEM: Improving LLMs' complex reasoning, especially commonsense, requires effective prompting strategies like chain-of-thought to break down problems and enhance factuality, though full understanding of their reasoning remains elusive.
    First, chain of thought, in principle, allows models to decompose multi-step problems into intermediate steps, which means that additional computation can be allocated to problems that require more reasoning steps.
    2. Second, a chain of thought provides an interpretable window into the behavior of the model, suggesting how it might have arrived at a particular answer and providing opportunities to debug where the reasoning path went wrong (although fully characterizing a model's computations that support an answer remains an open question).
  • RESEARCH_PROBLEM: The problem asks to calculate the total number of computers after additional installations over a period of time.
    Q: There were nine computers in the server room.
    Five more computers were installed each day, from monday to thursday.
  • RESEARCH_PROBLEM: The question asks how many pieces of pie were taken by the guests.
    How many pieces were taken by the guests?
  • RESEARCH_PROBLEM: The user has asked me to move the grapefruit drink to the counter.
    Explanation: The user has asked me to move the grapefruit drink to the counter.
  • RESEARCH_GAP: LLMs struggle with complex reasoning tasks like arithmetic and commonsense, with scaling alone insufficient; chain-of-thought prompting shows promise for improving performance without finetuning.
    However, scaling up model size alone has not proved sufficient for achieving high performance on challenging tasks such as arithmetic, commonsense, and symbolic reasoning (Rae et al., 2021).
    For the traditional fewshot prompting method used in Brown et al. (2020), it works poorly on tasks that require reasoning abilities, and often does not improve substantially with increasing language model scale (Rae et al., 2021).
  • RESEARCH_GAP: The coin flip task involves determining if a coin remains heads up after a series of flip or no-flip actions.
    This task asks the model to answer whether a coin is still heads up after people either flip or don't flip the coin (e.g., "A coin is heads up.
    Phoebe flips the coin.