generated at
SCoT
SCoT(Strategic Chain-of-Thought)
まず前提として、CoTと言う、llmの性能を引き出すprompt engeenring手法がある
AIによる論文の概要まとめ
> Chain-of-Thought (CoT) パラダイム: CoTは、LLMsの推論能力を強化するために、問題解決のための思考プロセスを明示的に生成するアプローチです。ただし、CoTは生成される推論パスの品質を一貫して保証できないため、不安定な結果をもたらすことがあります。
> Strategic Chain-of-Thought (SCoT): SCoTは、問題解決戦略を事前に明確にすることで、CoTの品質を向上させることを目的としています。具体的には、単一のプロンプト内で二段階のアプローチを採用します。まず、効果的な問題解決戦略を抽出します。その後、この戦略を基に高品質なCoTパスと最終的な答えを生成します
要するにゴールとその為の戦略を明確に書くことで解決策に迷いにくくするものらしい
Promptガイドライン
>SCoTプロンプトの構成要素
> Role (役割): モデルがどのような専門家として振る舞うべきかを定義します。
> Workflow (ワークフロー): 戦略の特定と適用のステップを概説します。
> Rules (ルール): 答えを生成するためのガイドラインを指定します。
> Initialization (初期化): タスクを設定します。
> Task Input (タスク入力): 解決すべき具体的な問題を提供します。
AIに書いてもらったサンプルPrompt(数学問題向け)
>Assume the role of a mathematician and solve the following problem using the most efficient strategy.
> Workflow:
> 1. Identify the most effective problem-solving strategy for this task.
> 2. Apply the identified strategy to generate a step-by-step reasoning path.
> 3. Use the reasoning path to derive the final answer.
> Rules:
> - Ensure each step is logically connected.
> - Validate intermediate results.
> Initialization:
> - Problem: (具体的な問題を記述)
> Task Input:
> - Given information: (与えられた情報)
> - Goal: (目標)
Automatic SCoT(SCoTを自動生成させる)
>Design a prompt template for Strategic Chain-of-Thought (SCoT) that integrates strategic knowledge before generating intermediate reasoning steps. The prompt should guide the model to identify the most effective problem-solving strategy and apply it to generate a high-quality chain-of-thought path.
> Workflow:
> 1. Identify the problem type and relevant strategic knowledge.
> 2. Generate a step-by-step reasoning path based on the identified strategy.
> 3. Ensure each step is logically connected and validate intermediate results.
> Example Problem: 具体的な問題を記述
> Task: Create a SCoT prompt template that can be used across various reasoning tasks.