Part 4 Advanced Zero-shot

In short: The model rephrases your question before answering it.

What is: Instruct the model to rewrite the question in its own words before answering. This clarifies intent and surfaces ambiguity.

When to use: Vague, ambiguous or poorly worded questions; when you want to make sure the model understood.

When to avoid: Already clear, direct questions, the rephrasing only adds text.

💡 Tip: Combine with CoT: rephrase, then reason. The clarity of the rephrasing improves the reasoning that follows.

In short: Ask the model to re-read the question before answering.

What is: Simply instruct the model to re-read the prompt, repeating the input. Cheap, improves comprehension on complex questions.

When to use: Long prompts or ones with many conditions, where details are easy to skip.

When to avoid: Short, simple questions.

💡 Tip: Near-zero-cost trick; worth stacking with CoT on problems with many constraints.

In short: The model asks itself sub-questions before the final answer.

What is: The model decides whether it needs follow-up sub-questions, answers each, and then composes the final answer. Ideal for questions that require combining several facts (multi-hop).

When to use: Questions depending on intermediate facts ("Who directed the film that won the Oscar the year X was born?").

When to avoid: Single-fact questions, nothing to decompose.

💡 Tip: Works even better with search access: each sub-question becomes a query.

In short: Clean the context of noise before answering.

What is: The model first rewrites the context removing irrelevant or distracting information, and only then answers on the clean context. Reduces the effect of noise and planted opinions.

When to use: Inputs with lots of noise, embedded opinion or conflicting information.

When to avoid: Already clean, objective inputs.

💡 Tip: Also useful as a defense: planted opinions and manipulation attempts are stripped in the cleaning step.

In short: The model generates analogous examples; or learns from right AND wrong.

What is: Analogical makes the model generate its own analogous examples before solving (self-few-shot). Contrastive CoT provides right AND wrong reasoning demonstrations so the model learns what to avoid.

When to use: Analogical: when you have no ready examples. Contrastive: when there's a recurring reasoning error to prevent.

When to avoid: Trivial tasks, generating analogues or contrasts is overhead.

💡 Tip: In Contrastive, show the EXACT error you want to avoid, the more specific, the more effective.

In short: CoT for long context; or a tree of logical explanations.

What is: Thread-of-Thought (ThoT) is a CoT adapted to long, chaotic contexts ("analyze in manageable parts"). Maieutic prompting builds a tree of recursive explanations and resolves contradictions logically.

When to use: ThoT: long, messy documents/histories. Maieutic: high-stakes logical reasoning.

When to avoid: Maieutic is expensive; reserve it for when logical consistency is essential.

💡 Tip: For very long contexts, combine ThoT with quote extraction before answering.