If a chatbot gives different answers each time you ask a factual question, how much should you trust any one of them? Self-consistency compares multiple answers to the same input. Agreement can suggest a more stable answer; disagreement can reveal that the model may be guessing. Neither establishes that an answer is true. section 5.7
To check for possible hallucinations, ask the same question through separate, direct calls to the model and compare the replies. Repeating it within one chatbot conversation is different: the chatbot may use its earlier replies as context and double down on a mistake. section 5.7
In self-consistency prompting, the model generates multiple reasoning chains and selects the most common answer. This extends chain-of-thought prompting, which follows one step-by-step path. Comparing reasoning paths may help with a problem, but selecting a frequent answer is not the same as independently checking the facts. section 7.4 section 5.7
A simple prompt
“Generate five solutions, then select the most common answer.” This asks for several routes to an answer rather than relying on the first one produced. section 7.4
Table 7.3 Common advanced prompting techniques
| Technique |
Description |
Example prompt |
|---|---|---|
| Chain-of-thought |
Generate step-by-step reasoning before answering |
“Let’s think step by step. John has 3 apples . . .” |
| Tree-of-thought |
Explore multiple reasoning paths in parallel |
“List three strategies, evaluate each, then choose the best.” |
| Self-consistency |
Generate multiple reasoning chains and aggregate |
“Generate five solutions, then select the most common answer.” |
| Reflexion |
Critique and revise output iteratively |
“Review your answer for errors and revise if needed.” |
| Verification |
Add a final self-check before finalizing the response |
“Double-check your answer for accuracy before submitting.” |
| Decomposition |
Break the problem into subtasks executed sequentially |
“Step 1: Extract metrics. Step 2: Summarize trends. Step 3: Write summary.” |
| Least-to-most |
Solve easier subproblems first, then build up |
“First, list known variables. Then calculate step by step.” |
| Pyramid priming |
Build knowledge gradually in layers |
“Step 1: Define carbon neutrality. Step 2: List targets. Step 3: Analyze impact.” |
| Generated knowledge |
Surface key facts before answering |
“List key concepts related to quantum entanglement before answering.” |
| Role-based prompting |
Assign a role or persona to guide behavior |
“You are an experienced employment lawyer advising a client.” |
| Prompt chaining |
Connect multiple prompts into a pipeline |
“Extract relevant quotes → Summarize → Generate final answer.” |
| Conversational context |
Summarize prior dialogue to preserve coherence |
“Summarize the student’s understanding before proceeding.” |
Compare self-consistency with other prompting approaches, including chain-of-thought, tree-of-thought, and verification. section 7.4
