Generated Knowledge Prompting
先 generate knowledge，再 answer（提升常识/知识密集型 tasks）
图片来源：Liu 等人 2022
LLM 继续得到改进，其中一种流行的技术是能够融合知识或信息，以帮助模型做出更准确的预测。
使用类似的思路，模型是否也可以在做出预测之前用于生成知识呢？这就是 Liu 等人 2022 的论文所尝试的——生成知识以作为提示的一部分。特别是，这对于常识推理等任务有多大帮助？
让我们尝试一个简单的提示：
提示：
Part of playing golf is trying to get a higher score than other people. Yes or no?
输出：
Yes.
这种错误揭示了 LLM 在执行需要更多关于世界的知识的任务时的局限性。我们如何通过生成知识来改进呢？
首先，我们生成一些“知识”：
提示：
Input: Greece is larger than Mexico.
Knowledge: Greece has an area of about 131,957 square kilometers, while Mexico has an area of about 1,964,375 square kilometers, making Mexico much larger than Greece.
Input: Glasses always fog up.
Knowledge: Glasses fog up when water vapor from your breath, sweat, or surrounding humidity condenses on a cooler lens surface, forming tiny droplets that look like a thin film.
Input: Fish can think.
Knowledge: Fish can show learning and memory abilities. Some fish demonstrate problem-solving and can retain information over time, which can help them navigate environments and social interactions.
Input: A common effect of smoking many cigarettes over a lifetime is a higher-than-normal chance of developing lung cancer.
Knowledge: Smoking increases the risk of lung cancer compared to never-smokers. The risk generally rises with greater smoking intensity and duration.
Input: A rock and a pebble are the same size.
Knowledge: In sedimentology (Udden-Wentworth scale), a pebble is typically 4 to 64 mm in diameter. Larger clasts (e.g., cobbles) are 64 to 256 mm, while smaller grains are below 4 mm.
Input: Part of playing golf is trying to get a higher score than other people.
Knowledge:
知识 1：
The goal of golf is to complete a set of holes in as few strokes as possible. A round typically includes 18 holes. Each stroke counts toward your total, and the player with the lowest total number of strokes wins.
知识 2：
Golf is a precision club-and-ball sport where players use various clubs to hit a ball into a series of holes on a course in as few strokes as possible. The objective is to finish with the lowest score, where score is the total number of strokes across all holes.
我们使用 Liu 等人 2022 的论文中提供的提示。
下一步是将知识整合并得出预测。我将问题重新格式化为 QA 格式，以指导答案格式。
提示：
Question: Part of playing golf is trying to get a higher score than other people. Yes or no?
Knowledge: The goal of golf is to complete a set of holes in as few strokes as possible. A round typically includes 18 holes. Each stroke counts toward your total, and the player with the lowest total number of strokes wins.
Explanation and answer:
答案 1（置信度非常高）：
No. In golf, the objective is not to get a higher score. Instead, players aim to take as few strokes as possible. The lowest total number of strokes determines the winner.
Question: Part of playing golf is trying to get a higher score than other people. Yes or no?
Knowledge: Golf is a precision club-and-ball sport where players use various clubs to hit a ball into a series of holes on a course in as few strokes as possible. The objective is to finish with the lowest score, where score is the total number of strokes across all holes.
Explanation and answer:
答案 2（置信度要低得多）：
No. In golf, players try to take the fewest strokes possible. The winner is determined by the lowest total number of strokes, not a higher score.
这个例子发生了一些非常有趣的事情。在第一个答案中，模型非常自信，但在第二个答案中不太自信。我简化了过程以进行演示，但在得出最终答案时还有一些细节需要考虑。请查看论文以了解更多。
📚 相关资源
❓ 常见问题
关于本章主题最常被搜索的问题，点击展开答案
Generated Knowledge prompting 是什么？
Liu 等人 2022 提出，让 LLM 在回答之前先「生成知识」作为 prompt 的一部分。两步走：第一步用 few-shot 让模型针对输入产出相关知识陈述；第二步把生成的知识 + 原问题重新喂给模型。专门用于常识推理任务，弥补模型常识盲区。
Generated Knowledge 和 RAG 是同一类东西吗？
不是。RAG 从外部数据源（向量库、维基）检索真实文档；Generated Knowledge 让 LLM 自己生成知识——本质是激活模型自己的参数化知识。前者抗幻觉强（信息可溯源），后者无需向量库但可能生成错误事实。两者也能结合：先 RAG 拉真实知识，再用模型扩写。
什么时候用 Generated Knowledge？
常识题、需要简单世界知识但不要求严格事实溯源的场景。论文示例：「打高尔夫的目标是分数比别人高吗？」直接问 LLM 答 Yes（错），先让它生成「高尔夫规则：杆数越少越好」再答就能纠正。也适合中小模型——大模型常识够强时收益变小。
Generated Knowledge 的「知识置信度」怎么影响答案？
原论文有个关键观察：同一个问题用两条不同的生成知识，模型给出的最终答案置信度差异明显——更精确、更聚焦的知识（如直接说「最少杆数获胜」）让模型答得更确信，模糊的知识让模型犹豫。最终决策需要结合多条知识聚合，而不是只用一条。
Generated Knowledge 怎么和 CoT 结合？
标准做法：先生成知识 → 拼接到 prompt → 让模型走 CoT 推理。比单纯 CoT 强在「知识层」被显式写出来，模型可以把它当作前提引用，避免凭空假设。也可以反过来——先 CoT 推理发现哪一步缺信息，再针对性生成那部分知识，做两轮迭代。