Side-by-Side Comparison
| Chain-of-Thought | Self-Consistency | Tree of Thoughts | |
|---|---|---|---|
| Shape of reasoning | One linear path | Several independent linear paths | Tree of partial paths with evaluation |
| How the answer is chosen | End of the single path | Majority vote across paths | Search: expand the best branches, backtrack from dead ends |
| Best for | Math, logic, multi-step analysis | Questions with one verifiable answer | Planning, puzzles, strategy, design trade-offs |
| Token cost | Low | Medium to high (N samples) | High (many branches and evaluations) |
| Runs in one prompt? | Yes | Yes, or N separate calls | Simplified version only; full version needs orchestration code |
| Introduced by | Wei et al., 2022 | Wang et al., 2022 | Yao et al., 2023 |
| Difficulty | Beginner | Intermediate | Advanced |
How Each Technique Works
Chain-of-Thought
You ask the model to show its intermediate steps before the final answer, either with worked examples (few-shot CoT) or a trigger such as "think step by step" (zero-shot CoT). Writing the steps out gives the model room to compute instead of guessing.
Self-Consistency
You generate several Chain-of-Thought answers with some randomness, then pick the final answer that appears most often. Different reasoning paths make different mistakes, so agreement between them is a useful signal of correctness.
Tree of Thoughts
You let the model propose several next steps, score them, and continue from the most promising ones, like a chess player considering candidate moves. When a branch looks hopeless, the search backtracks. This helps when the first plausible step is often the wrong one.
Example Prompts
Chain-of-Thought
A warehouse ships 1,240 orders on Monday and 18% more on Tuesday. On Wednesday it ships 150 fewer than Tuesday. How many orders did it ship in total? Think step by step, then give the final number on a line starting with "Answer:".
Self-Consistency
Solve the problem below three times using different reasoning. Label them Path 1–3 and end each with "Result:". Then report the result that appears most often and note any disagreement. [problem]
Tree of Thoughts (single-prompt version)
We need to cut cloud costs by 30% in two quarters without slowing releases. 1. Propose four distinct strategies. 2. Rate each for savings, risk and effort (1–10) and explain the scores. 3. Drop the two weakest. Break the remaining two into quarter-by-quarter steps. 4. If a step looks infeasible, backtrack to the next strategy. Output the final plan and why it beat the alternatives.
Decision Guide
- Everyday reasoning with a non-reasoning model? Chain-of-Thought.
- One correct answer and errors are costly? Self-Consistency with 3–5 samples.
- Planning or search where early choices matter? Tree of Thoughts.
- Using a reasoning model? Skip explicit CoT; consider Self-Consistency only for high-stakes answers.
Frequently Asked Questions
What is the difference between Chain-of-Thought and Tree of Thoughts?
Chain-of-Thought asks the model to reason along one linear path, step by step, to a single answer. Tree of Thoughts has the model propose several candidate reasoning steps at each stage, evaluate them, and keep exploring only the most promising branches, backtracking when a branch fails. CoT is one pass; ToT is a search.
How is Self-Consistency different from Tree of Thoughts?
Self-Consistency samples several complete, independent Chain-of-Thought answers and takes a majority vote on the final result. Tree of Thoughts evaluates partial reasoning while it is being built and can abandon or expand branches. Self-Consistency improves reliability for problems with one correct answer; Tree of Thoughts helps with planning and search problems where early choices matter.
Which technique uses the most tokens?
Tree of Thoughts is usually the most expensive because it generates and evaluates several branches at each step, often across multiple calls. Self-Consistency costs roughly the price of one Chain-of-Thought answer multiplied by the number of samples. Plain Chain-of-Thought is the cheapest of the three.
Do I still need Chain-of-Thought with reasoning models?
Much less. Reasoning models from OpenAI, Anthropic and Google already think step by step before answering, so adding 'think step by step' rarely helps and can make answers longer. Self-Consistency and Tree of Thoughts can still add value on top of reasoning models for high-stakes or search-heavy problems.
Can I run Tree of Thoughts in a normal chat interface?
A simplified version, yes: ask the model to propose several approaches, rate each one, then expand the best and backtrack if it fails. The full technique from the original paper runs as a program that calls the model many times and keeps the search state outside the model.