🎯 场景：Step-Back 技术——先"退一步"问出更抽象的元问题（"这个问题属于哪类"、"涉及什么原理"），用元问题的回答作为推理脚手架，再回到原题。降低被表面细节带偏的概率，对原理性 / 法律 / 物理 / 数学题特别有效。

Quick Use

Use when: A question's surface details might mislead direct reasoning, and reasoning from a more general principle would be more reliable. Fill in: {{question}} = the original question. You'll get: A more abstract "step-back" question, the answer to it, and the answer to the original question grounded in the principle. Output is JSON.

Purpose

Apply Step-Back Prompting (Zheng et al.) — instead of jumping directly to the answer, first generate a higher-level question that abstracts away the question's surface specifics, answer that abstract question, then use the abstract answer as a principle to ground the answer to the original question. Used on tasks where direct reasoning is prone to surface-feature errors: physics word problems, legal interpretation, deep-domain trivia, multi-condition queries. Output is structured so the abstraction step and the final answer are independently inspectable.

Prompt

You solve a question using "step-back" prompting: first abstract the
question into a more general principle question, answer the principle
question, then return to the specific question grounded in that
principle.

Original question:
{{question}}

Steps:
1. Generate a "step-back" question that:
   - Removes the specific entities, numbers, or constraints in the
     original.
   - Asks about the underlying principle, category, or rule that
     governs questions of this type.
   - Should be answerable in 1-2 sentences from general knowledge.
2. Answer the step-back question concisely (1-3 sentences). State
   the principle / rule / category.
3. Apply the principle to the original question to derive the
   final answer. Show the application explicitly: "By the principle
   that X, applied to {original question's specifics}, the answer is Y."

If the question doesn't benefit from abstraction (e.g. it's a direct
lookup, or the question already names the principle), set
"step_back_useful": false and answer directly in `final_answer`.

Return ONLY this JSON object:
{
  "step_back_useful": true | false,
  "step_back_question": "<the abstract question, or empty if step_back_useful=false>",
  "step_back_answer": "<the principle / rule / category, 1-3 sentences>",
  "application": "<one or two sentences explicitly applying the principle to the original question, or empty if step_back_useful=false>",
  "final_answer": "<the answer to the original question>",
  "decision_basis": "<one sentence, <=30 words, no internal CoT>"
}

Example

Input:

question: "If I drop a 5kg bowling ball and a 0.5kg apple from a 3-meter cliff at the same time, in a vacuum, which hits the ground first?"

Expected output:

{
  "step_back_useful": true,
  "step_back_question": "In a vacuum, does mass affect how fast an object falls under gravity?",
  "step_back_answer": "No. In a vacuum, all objects accelerate at the same rate (g ≈ 9.8 m/s² on Earth) regardless of mass — Galileo's principle of equivalence. Mass affects gravitational force but also inertia, and the two cancel.",
  "application": "Since both the bowling ball and the apple fall in a vacuum and start at the same height, they accelerate identically and reach the ground at the same time.",
  "final_answer": "They hit the ground at the same time.",
  "decision_basis": "The surface details (5kg vs 0.5kg) tempt a 'heavier falls faster' answer; stepping back to the equivalence principle prevents that error."
}

Failure Modes

• Cosmetic step-back — the "step-back" question is just the original with names removed, not actually more abstract. Detect by checking that step_back_question contains domain-general words (principle, rule, category, mechanism, concept) more often than domain-specific entities.
• Wrong principle — model abstracts to the wrong governing principle (e.g. invokes momentum conservation for a kinematics problem). Sample outputs and check application correctness.
• Step-back that loses essential context — abstraction discards the very thing that makes the question non-trivial (e.g. drops "in a vacuum" qualifier and then gives the air-resistance answer). Mitigation: rule "removes surface specifics" should preserve conditions that change which principle applies.
• Forced step-back — model insists on stepping back when the question is a direct lookup ("What's the boiling point of water at sea level?"). The step_back_useful: false path exists for this; if the rate is <5%, the bar for "useful" is too lax.
• Application gap — step_back_answer is correct but application doesn't actually use the principle to derive the answer. Audit outputs where final_answer is correct but application is empty or generic.
• Confidence collapse on hard questions — for genuinely difficult questions, model produces a vague step-back-answer to avoid committing. The principle should be a concrete claim; reject vague ones.

Tuning Notes

• 模型差异：本卡对模型的"识别问题类别"能力要求高。frontier 模型在 step_back 抽象的精准度上明显更稳；中档模型容易产生 cosmetic step-back。
• 温度：0.0–0.3。abstraction 必须可重现。
• 适用问题类型（按 ROI）：
  • 高：物理 / 数学带表面诱导（"重的物体下落更快"陷阱）、法律 场景（"违反规定"vs"违反原则"）、医学（症状与原理映射）
  • 中：deep-domain trivia（"哪个朝代的什么事件"——先识别朝代特征）
  • 低：直接 lookup（boiling point of water at sea level）
  • 不适合：纯创造性写作（步入抽象反而限制）
• 与 cot/least-to-most-decomposition 的对比：least-to-most 把问题 拆成更简单的子问题序列，是水平分解；step-back 把问题升维成 更抽象的问题，是垂直抽象。前者适合 compositional 任务，后者 适合原理性任务。
• 与 cot/structured-reasoning-with-rationale-summary 的关系： 本卡是更"结构化"的特殊形式——sub_steps 退化成 step_back_question + application 两步。可以认为 step-back 是 structured-reasoning 的 专用变种。
• 与 cot/self-consistency-aggregator 的关系：step-back 减少单路径 方差（通过抽象去 noise）；self-consistency 通过多采样投票去 noise。 两者正交，可叠加（多次跑 step-back 再投票）。
• 用作训练数据：(question, step_back_q, step_back_a, application, final_answer) 五元组可作为 Step-Back SFT 训练数据；建议手工 spot check 抽象质量后再入训练集。

Changelog

• 0.1.0 — Initial card.