Chain-of-Thought Prompting

CoT prompting forces the model to break a complex problem into a sequence of intermediate reasoning steps rather than jumping directly to a final answer. This is critical for tasks requiring multi-hop reasoning, where information from multiple sources must be synthesized.

• Transforms implicit cognition into an explicit, auditable reasoning trace
• Enables the model to allocate more inference-time compute to difficult sub-problems
• Example: For a math word problem, the model first extracts quantities, then identifies the operation, then computes, then verifies

In Few-Shot CoT, the prompt includes complete examples of questions paired with detailed reasoning chains. The model learns the pattern of step-by-step deliberation from these demonstrations without any gradient updates.

• The reasoning format is learned entirely in-context
• Effective even with a small number of carefully crafted exemplars
• Contrast with Zero-Shot CoT, which uses a simple trigger phrase like 'Let's think step by step' without examples

By externalizing the reasoning process, CoT reduces the likelihood of post-hoc rationalization and the Clever Hans Effect, where models rely on spurious correlations. The explicit chain makes it harder for the model to skip logical steps.

• Reduces hallucination snowballing by exposing each inferential leap to scrutiny
• Enables human auditors to identify exactly where a reasoning error occurred
• The trace serves as a faithfulness metric baseline for evaluating whether the output reflects true computation

CoT is often paired with Self-Consistency, a decoding strategy that samples multiple diverse reasoning paths and selects the most frequent final answer. This compensates for the stochasticity in individual chain generation.

• Improves reliability on tasks with a single correct answer
• Exploits the fact that there are often multiple valid ways to reason to the same conclusion
• The variance across sampled chains provides a crude uncertainty quantification signal

Standard linear CoT is the conceptual foundation for more sophisticated reasoning architectures. Tree-of-Thoughts extends it by exploring multiple branches simultaneously, while Graph-of-Thoughts allows merging and cycling of intermediate steps.

• Enables lookahead and backtracking during problem-solving
• Supports tool-augmented reasoning where steps involve external API calls
• The linear chain serves as the atomic unit that these frameworks compose and recombine

In enterprise contexts governed by the EU AI Act and similar regulations, CoT provides a form of algorithmic explainability. The reasoning trace serves as an auditable artifact demonstrating how a high-stakes decision was reached.

• Supports process supervision by allowing human reviewers to validate each step
• Contrasts with black-box outputs that offer no justification
• Enables compliance officers to verify that prohibited factors were not used in reasoning

A variant that elicits reasoning without any prior examples. By simply appending a trigger phrase like 'Let's think step by step' to a prompt, the model is nudged to generate an intermediate reasoning trace before producing the final answer. This technique is powerful because it requires no manual crafting of demonstrations, making it a universal, task-agnostic method for improving performance on arithmetic, symbolic, and commonsense reasoning tasks.

A decoding strategy that replaces the standard greedy approach. Instead of taking the single most probable reasoning path, it samples multiple diverse chains of thought for the same problem. The final answer is selected by a majority vote or by choosing the most consistent conclusion across all generated paths. This significantly improves reliability on tasks with multiple valid reasoning approaches by marginalizing over the model's stochastic generation process.

A reasoning trace that accurately reflects the true causal process by which the model arrived at its answer. This is distinct from a merely plausible or fluent explanation. A faithful chain-of-thought is free from post-hoc rationalization, where a model generates a convincing but inaccurate justification after the fact. Ensuring faithfulness is a core challenge in AI safety, as it determines whether a human can truly audit and trust the model's logic.

An extension that moves beyond linear reasoning by exploring multiple paths simultaneously in a tree structure. At each step, the model generates several potential 'thoughts', evaluates their promise via a state evaluator, and uses search algorithms like BFS or DFS to look ahead and backtrack. This deliberate planning mechanism allows the model to solve complex problems that require global exploration and lookahead, such as creative writing or puzzle-solving.

A training methodology that provides feedback on each intermediate step of a reasoning chain, not just the final outcome. A Process Reward Model (PRM) is trained to score the correctness of every logical increment. This contrasts with outcome supervision, which only verifies the final answer. Process supervision is crucial for training models to produce robust, aligned reasoning and for detecting subtle logical errors that happen to lead to a correct final result by chance.

A prompting paradigm that interleaves Reasoning traces and Action steps. The model generates a thought to analyze the situation, then an action to query an external tool (like a search API), and receives an observation to integrate into its next reasoning cycle. This synergistic loop grounds the chain-of-thought in external knowledge, significantly reducing hallucination on knowledge-intensive tasks like fact-checking and multi-hop question answering.