How to design prompts to test an AI's reasoning ability on complex problems?

When testing an AI's ability to reason through complex problems, prompt design should focus on constructing scenarios that require multi-step logical deduction, cross-domain information integration, or counterintuitive analysis, ensuring the problem contains clear constraints and hidden correlations. Multi-step logical scenarios: Design tasks with 2-3 interconnected sub-questions, such as "If Company A's 2023 revenue grew by 20% (reaching 10 million yuan), and in 2024, the cost-to-revenue ratio decreased from 30% to 25%, with the net profit margin increasing by 5 percentage points, calculate the 2024 net profit." This requires the AI to step-by-step decompose the growth base, cost changes, and the relationship with profit margins. Information conflict scenarios: Provide questions with contradictory data, for example, "According to Report A, a product's user retention rate is 70%; Report B shows that the churn rate during the same period is 40%. Please analyze possible reasons for the data discrepancy and verify its rationality," to test the AI's critical handling of conflicting information. Implicit premise mining: Pose questions that require supplementing unstated conditions, such as "Why is it difficult to cook food even when water boils at high altitudes?" which requires the AI to identify the implicit scientific principle that "boiling point decreases as atmospheric pressure drops." It is recommended to first clarify reasoning goals (e.g., causal analysis, data verification), gradually increase problem complexity, and compare the reasoning path differences of different AI models. This can help systematically evaluate the AI's logical chain completeness and vulnerability identification ability for complex problems.