Plan And Solve: Definition and Examples

Plan And Solve (PS) is a prompting strategy introduced by Wang et al. in 2023, designed to improve the ability of large language models to solve complex problems, especially in mathematics and logical reasoning. Instead of asking the model to solve a problem directly, it is explicitly asked to design an action plan before proceeding to execution.

This approach is inspired by the observation that methods like Chain of Thought (CoT) sometimes suffer from calculation errors, missing steps, or loss of coherence during reasoning. By forcing an explicit planning phase, Plan And Solve reduces these issues by structuring the model's thinking in two stages: first understand the problem and identify necessary steps, then execute each step in an orderly manner.

The improved variant, called PS+ (Plan And Solve Plus), adds additional instructions such as "extract variables and their values" or "compute intermediate results." These detailed guidelines guide the model further and significantly reduce reasoning errors.

Plan And Solve works in zero-shot mode, meaning it requires no prior examples in the prompt. Simply adding an instruction like "First devise a plan, then solve the problem step by step" yields significantly better results than a basic prompt.