Plan-and-Solve - Plan and Solve

LLM Zero-shot Reasoning Prompt Strategy P APIs & Messaging

Basic Information

Paradigm Description

The Plan-and-Solve (PS) prompting strategy improves LLM's zero-shot reasoning ability through two stages: first, devising a plan to break down the overall task into smaller subtasks, and then executing the subtasks step by step according to the plan. This method addresses the three major issues of Zero-shot CoT (calculation errors, missing steps, semantic misunderstandings) by introducing explicit planning steps to reduce missing step errors.

Core Mechanism

  • Two-stage approach: First understand the problem and devise a plan, then solve the problem step by step according to the plan
  • PS prompt format: "Let's first understand the problem and devise a plan to solve the problem. Then, let's carry out the plan and solve the problem step by step."
  • PS+ enhanced version: Adds detailed instructions such as "extract relevant variables and their values" and "calculate intermediate results (pay attention to calculations and common sense)"
  • QDMR extension: A new method based on Question Decomposition Meaning Representation, generating a directed acyclic graph to represent problem-solving logic

Performance

  • Consistently outperforms Zero-shot CoT on all datasets, with significant advantages
  • Comparable or superior to Zero-shot Program-of-Thought prompting method
  • Matches the performance of 8-shot CoT in mathematical reasoning
  • Effectively reduces missing step errors

Applicable Scenarios

  • Mathematical reasoning problems
  • Commonsense reasoning tasks
  • Symbolic reasoning tasks
  • Complex problems requiring multi-step planning

Limitations

  • Sensitive to prompt wording
  • Cannot completely resolve semantic misunderstanding errors
  • Relies on the LLM's basic reasoning ability

Relationship with Other Paradigms

  • vs Zero-shot CoT: PS explicitly introduces planning steps, reducing missing steps
  • vs Few-shot CoT: PS does not require manually written examples, it is a zero-shot method
  • vs ReAct: ReAct focuses on reasoning and environment interaction, PS focuses on planning reasoning steps

Relationship with the OpenClaw Ecosystem

The Plan-and-Solve paradigm can be applied to the task planning phase of OpenClaw agents. When users assign complex tasks to agents, the agents can first devise a plan, break down subtasks, and then execute them step by step according to the plan. This explicit planning step helps reduce step omissions when agents perform complex tasks, improving task completion rates.