Self-Ask

Multi-hop reasoning — answering a question that requires chaining multiple intermediate facts — is a known weakness of direct LLM prompting. Ask "What is the market capitalization of the company that acquired Slack?" and a model may produce a plausible-sounding but incorrect answer by conflating acquired companies, misremembering the acquirer, or using stale data. Self-Ask, introduced by Press et al. in 2022, addresses this by prompting the model to explicitly surface each intermediate question it needs to answer before tackling the original query.

The Self-Ask pattern works through few-shot exemplars that demonstrate the decomposition format. When the model encounters a complex question, it generates "Follow-up: [sub-question]" and "Intermediate answer: [answer]" pairs for each reasoning step, making the chain of logic fully explicit and auditable before arriving at "Final answer: [answer]". Crucially, each follow-up question can be directed to an external search or retrieval system, transforming Self-Ask into a structured retrieval orchestration pattern where the model itself determines what information it needs and when.

The enterprise value is highest in knowledge work that requires combining information from multiple sources: investment research that compares financial metrics across companies and time periods, legal analysis that applies regulatory rules to specific factual scenarios, technical due diligence that traces dependencies across a technology stack, and supply chain analysis that reasons across geography, logistics, and demand data. Self-Ask also produces naturally explainable outputs — the chain of follow-up questions and intermediate answers is an audit trail of the model's reasoning that compliance teams and business stakeholders can review and validate.