Chain-of-Thought Prompting (2022) › The Solution

Chain-of-thought prompting addressed these limitations by providing models with examples that showed explicit step-by-step reasoning before reaching final answers. Instead of prompt examples that went directly from problem to answer, chain-of-thought examples included intermediate reasoning steps that broke down complex problems into manageable subproblems. This simple modification activated latent reasoning capabilities in large language models, dramatically improving performance on complex tasks.

The core technique involved crafting few-shot examples where each example demonstrated a reasoning chain. For instance, rather than showing "Problem: X, Answer: Y," chain-of-thought prompting showed "Problem: X, Reasoning: Step 1..., Step 2..., Step 3..., Answer: Y." By seeing multiple examples of this reasoning pattern, models learned to generate intermediate steps before producing final answers, even for problems not seen in the examples.

The approach worked by teaching models to decompose problems systematically. When a model encountered a new problem after seeing chain-of-thought examples, it would generate intermediate reasoning steps similar to those in the examples, breaking the problem down before arriving at an answer. This decomposition process activated the model's ability to reason through sequential steps, capabilities that existed in the training data but weren't being elicited by standard prompting.

The formulation used phrases like "Let's think step by step" or "Let's work through this" to explicitly signal that reasoning chains were desired. These simple cues, combined with example reasoning chains, taught models to produce intermediate steps rather than jumping directly to answers. The key insight was that models could learn to reason explicitly when prompted appropriately, even though their training didn't emphasize reasoning chains as a special capability.

Chain-of-thought prompting proved particularly effective for arithmetic reasoning tasks. Mathematical word problems that required multiple operations benefited enormously from step-by-step decomposition. Models could now solve problems like "If Alice has 3 apples and gives 2 to Bob, then receives 5 more, how many does she have?" by generating steps: "Alice starts with 3 apples. After giving 2 to Bob, she has $3 - 2 = 1$ apple. After receiving 5 more, she has $1 + 5 = 6$ apples." This explicit reasoning dramatically improved accuracy on complex arithmetic problems.

The technique also improved performance on symbolic reasoning tasks. Logical puzzles, analogy problems, and sequence completion tasks all benefited from explicit reasoning chains. Models could now work through logical deductions step by step, considering intermediate conclusions before reaching final answers. This capability was particularly valuable for problems where the reasoning path mattered as much as the final answer.

An important variant of the approach involved zero-shot chain-of-thought prompting, where models were simply prompted with phrases like "Let's think step by step" without providing examples. This variant showed that models could generate reasoning chains even without explicit examples, though few-shot examples generally produced better results. The zero-shot capability suggested that reasoning was a fundamental capability that just needed to be activated, not taught from scratch.

The approach also enabled better handling of multi-step problems where intermediate results needed to be computed and used in subsequent steps. Problems requiring planning, scheduling, or sequential decision making could be broken down into explicit steps, allowing models to track intermediate states and make decisions at each stage. This capability made models much more effective at tasks requiring sequential reasoning.