作者：紫氣東來(lái)

項(xiàng)目地址：https://zhuanlan.zhihu.com/p/632369186

一、概述

提示工程（Prompt Engineering），也稱為 In-Context Prompting，是指在不更新模型權(quán)重的情況下如何與 LLM 交互以引導(dǎo)其行為以獲得所需結(jié)果的方法。 在提示工程中，任務(wù)的描述會(huì)被嵌入到輸入中。例如，不是隱含地給予模型一定的參數(shù)，而是以問(wèn)題的形式直接輸入。 提示工程的典型工作方式是將一個(gè)或多個(gè)任務(wù)轉(zhuǎn)換為基于提示的數(shù)據(jù)集，并通過(guò)所謂的“基于提示的學(xué)習(xí)（prompt-based learning）”來(lái)訓(xùn)練語(yǔ)言模型。

提示工程不僅僅是關(guān)于設(shè)計(jì)和研發(fā)提示詞。它包含了與大語(yǔ)言模型交互和研發(fā)的各種技能和技術(shù)。提示工程在實(shí)現(xiàn)和大語(yǔ)言模型交互、對(duì)接，以及理解大語(yǔ)言模型能力方面都起著重要作用。用戶可以通過(guò)提示工程來(lái)提高大語(yǔ)言模型的安全性，也可以賦能大語(yǔ)言模型，比如借助專業(yè)領(lǐng)域知識(shí)和外部工具來(lái)增強(qiáng)大語(yǔ)言模型能力。

提示詞可以包含以下任意要素：

• 指令：想要模型執(zhí)行的特定任務(wù)或指令。

• 上下文：包含外部信息或額外的上下文信息，引導(dǎo)語(yǔ)言模型更好地響應(yīng)。

• 輸入數(shù)據(jù)：用戶輸入的內(nèi)容或問(wèn)題。

• 輸出指示：指定輸出的類型或格式。

以下是設(shè)計(jì)提示的通用技巧:

• 從簡(jiǎn)單開始：在設(shè)計(jì)提示時(shí)，需要記住這是一個(gè)迭代的過(guò)程，需要大量的實(shí)驗(yàn)來(lái)獲得最佳結(jié)果?？梢詮暮?jiǎn)單的提示開始，不斷添加更多的元素和上下文，以獲得更好的結(jié)果。

• 指令：可以使用命令來(lái)指示模型執(zhí)行各種簡(jiǎn)單任務(wù)，例如“寫入”、“分類”、“總結(jié)”、“翻譯”、“排序”等，從而為各種簡(jiǎn)單任務(wù)設(shè)計(jì)有效的提示。

• 具體性：對(duì)希望模型執(zhí)行的指令和任務(wù)，提示越具體和詳細(xì)，結(jié)果就越好。實(shí)際上，在提示中提供示例非常有效，可以以特定格式獲得所需的輸出。

• 避免不精確：這里的類比非常類似于有效的溝通——越直接，信息傳遞就越有效。

• 做還是不做：設(shè)計(jì)提示時(shí)的另一個(gè)常見(jiàn)技巧是避免說(shuō)不要做什么，而是說(shuō)要做什么。

二、提示技術(shù)

時(shí)至今日，改進(jìn)提示顯然有助于在不同任務(wù)上獲得更好的結(jié)果。這就是提示工程背后的整個(gè)理念。在本節(jié)中，我們將介紹更高級(jí)的提示工程技術(shù)，使我們能夠完成更復(fù)雜和有趣的任務(wù)，所有測(cè)試案例均通過(guò)text-davinci-003 得到。

2.1 Zero-shot 與 Few-shot

Zero-shot 與 Few-shot 是最基礎(chǔ)的提示技術(shù)。經(jīng)過(guò)大量數(shù)據(jù)訓(xùn)練并調(diào)整指令的LLM能夠執(zhí)行 Zero-shot 任務(wù)，即直接向模型輸入文本以獲取回答。

如，Zero-shot 輸入：

Text: I'll bet the video game is a lot more fun than the film.Sentiment:

輸出：

Positive - The speaker expresses that they think the video game is more enjoyable than the film.

Few-shot learning 在目標(biāo)任務(wù)上提供了一組高質(zhì)量的演示，每個(gè)演示都包含輸入和期望的輸出。 當(dāng)模型首先看到好的例子時(shí)，它可以更好地理解人類的意圖和需要什么樣的答案的標(biāo)準(zhǔn)。 因此，少樣本學(xué)習(xí)通常比零樣本學(xué)習(xí)有更好的性能。 然而，它是以更多的 token 消耗為代價(jià)的，并且當(dāng)輸入和輸出文本很長(zhǎng)時(shí)可能會(huì)達(dá)到上下文長(zhǎng)度限制。

如，F(xiàn)ew-shot 輸入：

Text: (lawrence bounces) all over the stage, dancing, running, sweating, mopping his face and generally displaying the wacky talent that brought him fame in the first place.Sentiment: positiveText: despite all evidence to the contrary, this clunker has somehow managed to pose as an actual feature movie, the kind that charges full admission and gets hyped on tv and purports to amuse small children and ostensible adults.Sentiment: negativeText: for the first time in years, de niro digs deep emotionally, perhaps because he's been stirred by the powerful work of his co-stars.Sentiment: positiveText: I'll bet the video game is a lot more fun than the film.Sentiment:

輸出：

negative

2.2 思維鏈（Chain-of-Thought, CoT）提示

CoT 提示 生成一系列短句來(lái)逐步描述推理邏輯，稱為推理鏈或基本原理，最終得出最終答案。 CoT 的好處對(duì)于復(fù)雜的推理任務(wù)更為，同時(shí)使用大型模型（例如，參數(shù)超過(guò) 50B）時(shí)效果更加明顯。






如，F(xiàn)ew-shot CoT 輸入：

Question: Tom and Elizabeth have a competition to climb a hill. Elizabeth takes 30 minutes to climb the hill. Tom takes four times as long as Elizabeth does to climb the hill. How many hours does it take Tom to climb up the hill?Answer: It takes Tom 30*4 = <<30*4=120>>120 minutes to climb the hill.It takes Tom 120/60 = <<120/60=2>>2 hours to climb the hill.So the answer is 2.===Question: Jack is a soccer player. He needs to buy two pairs of socks and a pair of soccer shoes. Each pair of socks cost $9.50, and the shoes cost $92. Jack has $40. How much more money does Jack need?Answer: The total cost of two pairs of socks is $9.50 x 2 = $<<9.5*2=19>>19.The total cost of the socks and the shoes is $19 + $92 = $<<19+92=111>>111.Jack need $111 - $40 = $<<111-40=71>>71 more.So the answer is 71.===Question: Marty has 100 centimeters of ribbon that he must cut into 4 equal parts. Each of the cut parts must be divided into 5 equal parts. How long will each final cut be?Answer:

輸出：

Marty has 100 cm of ribbon that needs to be cut into 4 equal parts.Each of the four cuts will be 100/4 = <<100/4=25>>25 cm long.Marty must then divide each of the four cuts into 5 equal parts. Each of the five final cuts will be 25/5 = <<25/5=5>>5 cm long.So the answer is 5 cm.

如，Zero-shot CoT 輸入：

Question: Marty has 100 centimeters of ribbon that he must cut into 4 equal parts. Each of the cut parts must be divided into 5 equal parts. How long will each final cut be?Answer: Let's think step by step.

輸出：

Step 1: Marty has to cut 100 centimeters of ribbon into 4 equal parts. So, each of the parts will be 25 centimeters long.Step 2: Marty has to divide each of the 4 parts into 5 equal parts. So, each of the final cut parts will be 5 centimeters long.

2.3 指令提示（Instruction Prompting）

Instructed LM（例如 InstructGPT）使用高質(zhì)量數(shù)據(jù)微調(diào)預(yù)訓(xùn)練模型，使 LM 更好地理解用戶意圖并遵循指令。 RLHF 是一種常用的方法。 instruction style fine-tuning 的好處是改進(jìn)了模型，使其更符合人類的意圖，并大大降低了溝通成本。

在與指令模型交互時(shí)，我們應(yīng)該詳細(xì)描述任務(wù)要求，盡量具體和準(zhǔn)確，避免說(shuō)“不做某事”，而是具體說(shuō)明要做什么。如針對(duì)特定人群的輸入：

Describe what is quantum physics to a 6-year-old.

輸出：

Quantum physics is like a super special game. It's all about understanding the very tiny things that make up our world. It's like trying to find out how puzzle pieces fit together to make a big picture.

In-context instruction learning 將小樣本學(xué)習(xí)與指令提示相結(jié)合。 它在提示中包含多個(gè)跨不同任務(wù)的演示示例，每個(gè)演示都由指令、任務(wù)輸入和輸出組成。 請(qǐng)注意，他們的實(shí)驗(yàn)僅針對(duì)分類任務(wù)，指令提示包含所有標(biāo)簽選項(xiàng)。

如，In-context instruction learning 輸入：

Definition: Determine the speaker of the dialogue, "agent" or "customer".Input: I have successfully booked your tickets.Ouput: agentDefinition: Determine which category the question asks for, "Quantity" or "Location".Input: What's the oldest building in US?Ouput: LocationDefinition: Classify the sentiment of the given movie review, "positive" or "negative".Input: i'll bet the video game is a lot more fun than the film.Output:

輸出：

Positive

2.4 自我一致性（Self-Consistency）采樣

自我一致性旨在“替換鏈?zhǔn)剿季S提示中使用的天真貪婪解碼方法”。其想法是通過(guò)少樣本CoT采樣多個(gè)不同的推理路徑，并使用生成結(jié)果選擇最一致的答案。這有助于提高CoT提示在涉及算術(shù)和常識(shí)推理的任務(wù)中的性能。

嘗試進(jìn)行以下數(shù)學(xué)推理問(wèn)題：

When I was 6 my sister was half my age. Now I’m 70 how old is my sister?

結(jié)果如下：

Your sister is now 35 years old.

結(jié)果錯(cuò)誤，下面使用自我一致性來(lái)嘗試該問(wèn)題，輸入如下：

Q: There are 15 trees in the grove. Grove workers will plant trees in the grove today. After they are done,there will be 21 trees. How many trees did the grove workers plant today?A: We start with 15 trees. Later we have 21 trees. The difference must be the number of trees they planted.So, they must have planted 21 - 15 = 6 trees. The answer is 6.Q: If there are 3 cars in the parking lot and 2 more cars arrive, how many cars are in the parking lot?A: There are 3 cars in the parking lot already. 2 more arrive. Now there are 3 + 2 = 5 cars. The answer is 5.Q: Leah had 32 chocolates and her sister had 42. If they ate 35, how many pieces do they have left in total?A: Leah had 32 chocolates and Leah’s sister had 42. That means there were originally 32 + 42 = 74chocolates. 35 have been eaten. So in total they still have 74 - 35 = 39 chocolates. The answer is 39.Q: Jason had 20 lollipops. He gave Denny some lollipops. Now Jason has 12 lollipops. How many lollipopsdid Jason give to Denny?A: Jason had 20 lollipops. Since he only has 12 now, he must have given the rest to Denny. The number oflollipops he has given to Denny must have been 20 - 12 = 8 lollipops. The answer is 8.Q: Shawn has five toys. For Christmas, he got two toys each from his mom and dad. How many toys doeshe have now?A: He has 5 toys. He got 2 from mom, so after that he has 5 + 2 = 7 toys. Then he got 2 more from dad, soin total he has 7 + 2 = 9 toys. The answer is 9.Q: There were nine computers in the server room. Five more computers were installed each day, frommonday to thursday. How many computers are now in the server room?A: There are 4 days from monday to thursday. 5 computers were added each day. That means in total 4 * 5 =20 computers were added. There were 9 computers in the beginning, so now there are 9 + 20 = 29 computers.The answer is 29.Q: Michael had 58 golf balls. On tuesday, he lost 23 golf balls. On wednesday, he lost 2 more. How manygolf balls did he have at the end of wednesday?A: Michael initially had 58 balls. He lost 23 on Tuesday, so after that he has 58 - 23 = 35 balls. OnWednesday he lost 2 more so now he has 35 - 2 = 33 balls. The answer is 33.Q: Olivia has $23. She bought five bagels for $3 each. How much money does she have left?A: She bought 5 bagels for $3 each. This means she spent 5Q: When I was 6 my sister was half my age. Now I’m 70 how old is my sister?A:

輸出1：

When you were 6, your sister was 6/2 = 3 years old. Now you are 70 years old. That means your sister is 70- 3 = 67 years old. The answer is 67.

輸出2:

When you were 6, your sister was 6 / 2 = 3 years old. Now that you are 70 years old, your sister mustbe 70 / 2 = 35 years old. The answer is 35.

輸出3:

When I was 6, my sister was 3 (half my age). 70 years have passed since then. That means my sister is now 3 + 70 = 73 years old. The answer is 73.

2.5 自動(dòng)提示工程師（Automatic Prompt Engineer, APE）

APE 是一種搜索模型生成的候選指令池，然后根據(jù)所選得分函數(shù)過(guò)濾候選集以最終選擇得分最高的最佳候選指令的方法。其過(guò)程可概括為3個(gè)階段：

1. 
   

提示 LLM 根據(jù)輸入輸出對(duì)形式的一小組演示生成候選指令。如：{{Given desired input-output pairs}}\n\nThe instruction is ；






1. 使用迭代蒙特卡洛搜索方法通過(guò)提示（如：Generate a variation of the following instruction while keeping the semantic meaning.\n\nInput: ...\n\nOutput:...）來(lái)提出語(yǔ)義相似的變體來(lái)改進(jìn)最佳候選者。

為了構(gòu)造自動(dòng)戶 CoT 提示，Shum et al. (2023) 建議進(jìn)行剪枝選擇，包括以下3步：

1. 增強(qiáng)(Augment)：使用 Few-shot 或 Zero-shot CoT 提示生成給定問(wèn)題的多個(gè)偽思維鏈；

2. 修剪(Prune)：根據(jù)生成的答案是否與基本事實(shí)相匹配來(lái)修剪偽鏈。

3. 選擇(Select)：應(yīng)用減少方差的策略梯度策略來(lái)學(xué)習(xí)所選示例的概率分布，同時(shí)將示例的概率分布視為策略，將驗(yàn)證集的準(zhǔn)確性視為獎(jiǎng)勵(lì)。

Zhang et al. (2023) 認(rèn)為采用聚類技術(shù)對(duì)問(wèn)題進(jìn)行抽樣，然后生成鏈。 他們觀察到 LLM 傾向于犯某些類型的錯(cuò)誤。 一種類型的錯(cuò)誤在嵌入空間中可能相似，因此被組合在一起。 通過(guò)僅從頻繁錯(cuò)誤的集群中抽取一個(gè)或幾個(gè)樣本，我們可以防止對(duì)一種錯(cuò)誤類型的過(guò)多錯(cuò)誤演示，并收集一組不同的示例。

1. 問(wèn)題聚類(Question clustering)：Embed 問(wèn)題使用 k-means 的方法進(jìn)行聚類。

2. 示例選擇(Demonstration selection)：從每個(gè)集群中選擇一組有代表性的問(wèn)題； 即來(lái)自一個(gè)集群的一個(gè)示例。 每個(gè)簇中的樣本按到簇質(zhì)心的距離排序，最接近質(zhì)心的樣本首先被選擇。

3. 論據(jù)生成(Rationale generation)：使用 Zero-shot CoT 為選定的問(wèn)題生成推理鏈，并構(gòu)建 Few-shot 提示以運(yùn)行推理。

三、更多資料

3.1 實(shí)用工具

• OpenAI Cookbook has many in-depth examples for how to utilize LLM efficiently.

• LangChain, a library for combining language models with other components to build applications.

• Prompt Engineering Guide repo contains a pretty comprehensive collection of education materials on prompt engineering.

• learnprompting.org

• PromptPerfect

• Semantic Kernel

3.2 數(shù)據(jù)集

• Anthropic's Red Team dataset(opens in a new tab)，(論文)(opens in a new tab)

• Awesome ChatGPT Prompts(opens in a new tab)

• DiffusionDB(opens in a new tab)

• Midjourney Prompts(opens in a new tab)

• P3 - Public Pool of Prompts(opens in a new tab)

• PartiPrompts(opens in a new tab)

• Real Toxicity Prompts(opens in a new tab)

• Stable Diffusion Dataset(opens in a new tab)

• WritingPrompts(opens in a new tab)

3.3 相關(guān)論文

綜述

• Nature Language Reasoning, A Survey(opens in a new tab) (March 2023)

• Augmented Language Models: a Survey(opens in a new tab) (Feb 2023)

• A Survey for In-context Learning(opens in a new tab) (Dec 2022)

• Towards Reasoning in Large Language Models: A Survey(opens in a new tab) (Dec 2022)

• Reasoning with Language Model Prompting: A Survey(opens in a new tab) (Dec 2022)

• Emergent Abilities of Large Language Models(opens in a new tab) (Jun 2022)

• A Taxonomy of Prompt Modifiers for Text-To-Image Generation(opens in a new tab) (Apr 2022)

• Pre-train, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing(opens in a new tab) (Jul 2021)

方法

• Self-Refine: Iterative Refinement with Self-Feedback(opens in a new tab) (Mar 2023)

• kNN Prompting: Beyond-Context Learning with Calibration-Free Nearest Neighbor Inference(opens in a new tab) (Mar 2023)

• Visual-Language Prompt Tuning with Knowledge-guided Context Optimization(opens in a new tab) (Mar 2023)

• Fairness-guided Few-shot Prompting for Large Language Models(opens in a new tab) (Mar 2023)

• Context-faithful Prompting for Large Language Models(opens in a new tab) (Mar 2023)

• Is Prompt All You Need? No. A Comprehensive and Broader View of Instruction Learning(opens in a new tab) (Mar 2023)

• UPRISE: Universal Prompt Retrieval for Improving Zero-Shot Evaluation(opens in a new tab) (Mar 2023)

• Model-tuning Via Prompts Makes NLP Models Adversarially Robust(opens in a new tab) (Mar 2023)

• Structure Pretraining and Prompt Tuning for Knowledge Graph Transfer(opens in a new tab) (March 2023)

• CoTEVer: Chain of Thought Prompting Annotation Toolkit for Explanation Verification(opens in a new tab) (March 2023)

• Larger language models do in-context learning differently(opens in a new tab) (March 2023)

• OpenICL: An Open-Source Framework for In-context Learning(opens in a new tab) (March 2023)

• Dynamic Prompting: A Unified Framework for Prompt Tuning(opens in a new tab) (March 2023)

• Multitask Prompt Tuning Enables Parameter-Efficient Transfer Learning(opens in a new tab) (March 2023)

• Effectiveness of Data Augmentation for Prefix Tuning with Limited Data(opens in a new tab) (March 2023)

• Mixture of Soft Prompts for Controllable Data Generation(opens in a new tab) (March 2023)

• Prompt, Generate, then Cache: Cascade of Foundation Models makes Strong Few-shot Learners(opens in a new tab) (March 2023)

• How Robust is GPT-3.5 to Predecessors? A Comprehensive Study on Language Understanding Tasks(opens in a new tab) (March 2023)

• Can ChatGPT Understand Too? A Comparative Study on ChatGPT and Fine-tuned BERT(opens in a new tab) (Feb 2023)

• EvoPrompting: Language Models for Code-Level Neural Architecture Search(opens in a new tab) (Feb 2023)

• In-Context Instruction Learning(opens in a new tab) (Feb 2023)

• Chain of Hindsight Aligns Language Models with Feedback(opens in a new tab) (Feb 2023)

• Language Is Not All You Need: Aligning Perception with Language Models(opens in a new tab) (Feb 2023)

• Automatic Prompt Augmentation and Selection with Chain-of-Thought from Labeled Data(opens in a new tab) (Feb 2023)

• Active Prompting with Chain-of-Thought for Large Language Models(opens in a new tab) (Feb 2023)

• More than you've asked for: A Comprehensive Analysis of Novel Prompt Injection Threats to Application-Integrated Large Language Models(opens in a new tab) (Feb 2023)

• A Prompt Pattern Catalog to Enhance Prompt Engineering with ChatGPT(opens in a new tab) (Feb 2023)

• Guiding Large Language Models via Directional Stimulus Prompting(opens in a new tab) (Feb 2023)

• How Does In-Context Learning Help Prompt Tuning?(opens in a new tab) (Feb 2023)

• Scalable Prompt Generation for Semi-supervised Learning with Language Models(opens in a new tab) (Feb 2023)

• Bounding the Capabilities of Large Language Models in Open Text Generation with Prompt Constraints(opens in a new tab) (Feb 2023)

• à-la-carte Prompt Tuning (APT): Combining Distinct Data Via Composable Prompting(opens in a new tab) (Feb 2023)

• GraphPrompt: Unifying Pre-Training and Downstream Tasks for Graph Neural Networks(opens in a new tab) (Feb 2023)

• The Capacity for Moral Self-Correction in Large Language Models(opens in a new tab) (Feb 2023)

• SwitchPrompt: Learning Domain-Specific Gated Soft Prompts for Classification in Low-Resource Domains(opens in a new tab) (Feb 2023)

• Evaluating the Robustness of Discrete Prompts(opens in a new tab) (Feb 2023)

• Compositional Exemplars for In-context Learning(opens in a new tab) (Feb 2023)

• Hard Prompts Made Easy: Gradient-Based Discrete Optimization for Prompt Tuning and Discovery(opens in a new tab) (Feb 2023)

• Multimodal Chain-of-Thought Reasoning in Language Models(opens in a new tab) (Feb 2023)

• Large Language Models Can Be Easily Distracted by Irrelevant Context(opens in a new tab) (Feb 2023)

• Synthetic Prompting: Generating Chain-of-Thought Demonstrations for Large Language Models(opens in a new tab) (Feb 2023)

• Progressive Prompts: Continual Learning for Language Models(opens in a new tab) (Jan 2023)

• Batch Prompting: Efficient Inference with LLM APIs(opens in a new tab) (Jan 2023)

• Demonstrate-Search-Predict: Composing retrieval and language models for knowledge-intensive NLP(opens in a new tab) (Dec 2022)

• On Second Thought, Let's Not Think Step by Step! Bias and Toxicity in Zero-Shot Reasoning(opens in a new tab) (Dec 2022)

• Constitutional AI: Harmlessness from AI Feedback(opens in a new tab) (Dec 2022)

• Successive Prompting for Decomposing Complex Questions(opens in a new tab) (Dec 2022)

• Large Language Models are reasoners with Self-Verification(opens in a new tab) (Dec 2022)

• Discovering Language Model Behaviors with Model-Written Evaluations(opens in a new tab) (Dec 2022)

• Structured Prompting: Scaling In-Context Learning to 1,000 Examples(opens in a new tab) (Dec 2022)

• PAL: Program-aided Language Models(opens in a new tab) (Nov 2022)

• Large Language Models Are Human-Level Prompt Engineers(opens in a new tab) (Nov 2022)

• Ignore Previous Prompt: Attack Techniques For Language Models(opens in a new tab) (Nov 2022)

• Machine Generated Text: A Comprehensive Survey of Threat Models and Detection Methods(opens in a new tab) (Nov 2022)

• Teaching Algorithmic Reasoning via In-context Learning(opens in a new tab) (Nov 2022)

• Enhancing Self-Consistency and Performance of Pre-Trained Language Models through Natural Language Inference(opens in a new tab) (Nov 2022)

• Ask Me Anything: A simple strategy for prompting language models(opens in a new tab) (Oct 2022)

• Recitation-Augmented Language Models(opens in a new tab) (Oct 2022)

• ReAct: Synergizing Reasoning and Acting in Language Models(opens in a new tab) (Oct 2022)

• Prompting GPT-3 To Be Reliable(opens in a new tab) (Oct 2022)

• Decomposed Prompting: A Modular Approach for Solving Complex Tasks(opens in a new tab) (Oct 2022)

• Language Models Are Greedy Reasoners: A Systematic Formal Analysis of Chain-of-Thought(opens in a new tab) (Oct 2022)

• Evaluating the Susceptibility of Pre-Trained Language Models via Handcrafted Adversarial Examples(opens in a new tab) (Sep 2022)

• Dynamic Prompt Learning via Policy Gradient for Semi-structured Mathematical Reasoning(opens in a new tab) (Sep 2022)

• Promptagator: Few-shot Dense Retrieval From 8 Examples(opens in a new tab) (Sep 2022)

• Atlas: Few-shot Learning with Retrieval Augmented Language Models(opens in a new tab) (Nov 2022)

• DocPrompting: Generating Code by Retrieving the Docs(opens in a new tab) (July 2022)

• On the Advance of Making Language Models Better Reasoners(opens in a new tab) (June 2022)

• Large Language Models are Zero-Shot Reasoners(opens in a new tab) (May 2022)

• Maieutic Prompting: Logically Consistent Reasoning with Recursive Explanations(opens in a new tab) (May 2022)

• MRKL Systems: A modular, neuro-symbolic architecture that combines large language models, external knowledge sources and discrete reasoning(opens in a new tab) (May 2022)

• PPT: Pre-trained Prompt Tuning for Few-shot Learning(opens in a new tab) (Mqy 2022)

• Toxicity Detection with Generative Prompt-based Inference(opens in a new tab) (May 2022)

• Learning to Transfer Prompts for Text Generation(opens in a new tab) (May 2022)

• The Unreliability of Explanations in Few-shot Prompting for Textual Reasoning(opens in a new tab) (May 2022)

• A Taxonomy of Prompt Modifiers for Text-To-Image Generation(opens in a new tab) (Apr 2022)

• PromptChainer: Chaining Large Language Model Prompts through Visual Programming(opens in a new tab) (Mar 2022)

• Self-Consistency Improves Chain of Thought Reasoning in Language Models(opens in a new tab) (March 2022)

• Training language models to follow instructions with human feedback(opens in a new tab)

• Rethinking the Role of Demonstrations: What Makes In-Context Learning Work?(opens in a new tab) (Feb 2022)

• Chain of Thought Prompting Elicits Reasoning in Large Language Models(opens in a new tab) (Jan 2022)

• Show Your Work: Scratchpads for Intermediate Computation with Language Models(opens in a new tab) (Nov 2021)

• AI Chains: Transparent and Controllable Human-AI Interaction by Chaining Large Language Model Prompts(opens in a new tab) (Oct 2021)

• Generated Knowledge Prompting for Commonsense Reasoning(opens in a new tab) (Oct 2021)

• Multitask Prompted Training Enables Zero-Shot Task Generalization(opens in a new tab) (Oct 2021)

• Reframing Instructional Prompts to GPTk's Language(opens in a new tab) (Sep 2021)

• Design Guidelines for Prompt Engineering Text-to-Image Generative Models(opens in a new tab) (Sep 2021)

• Making Pre-trained Language Models Better Few-shot Learners(opens in a new tab) (Aug 2021)

• Fantastically Ordered Prompts and Where to Find Them: Overcoming Few-Shot Prompt Order Sensitivity(opens in a new tab) (April 2021)

• BERTese: Learning to Speak to BERT(opens in a new tab) (April 2021)

• The Power of Scale for Parameter-Efficient Prompt Tuning(opens in a new tab) (April 2021)

• Prompt Programming for Large Language Models: Beyond the Few-Shot Paradigm(opens in a new tab) (Feb 2021)

• Calibrate Before Use: Improving Few-Shot Performance of Language Models(opens in a new tab) (Feb 2021)

• Prefix-Tuning: Optimizing Continuous Prompts for Generation(opens in a new tab) (Jan 2021)

• Learning to Generate Task-Specific Adapters from Task Description(opens in a new tab) (Jan 2021)

• Making Pre-trained Language Models Better Few-shot Learners(opens in a new tab) (Dec 2020)

• Learning from Task Descriptions(opens in a new tab) (Nov 2020)

• AutoPrompt: Eliciting Knowledge from Language Models with Automatically Generated Prompts(opens in a new tab) (Oct 2020)

• Language Models are Few-Shot Learners(opens in a new tab) (May 2020)

• How Can We Know What Language Models Know?(opens in a new tab) (July 2020)

• Scaling Laws for Neural Language Models(opens in a new tab) (Jan 2020)

應(yīng)用

• PaLM 2 Technical Report(opens in a new tab) (May 2023)

• BloombergGPT: A Large Language Model for Finance(opens in a new tab) (March 2023)

• Medical Intervention Duration Estimation Using Language-enhanced Transformer Encoder with Medical Prompts(opens in a new tab) (March 2023)

• Soft-prompt tuning to predict lung cancer using primary care free-text Dutch medical notes(opens in a new tab) (March 2023)

• TaskMatrix.AI: Completing Tasks by Connecting Foundation Models with Millions of APIs(opens in a new tab) (March 2023)

• Larger Probes Tell a Different Story: Extending Psycholinguistic Datasets Via In-Context Learning(opens in a new tab) (March 2023)

• Linguistically Informed ChatGPT Prompts to Enhance Japanese-Chinese Machine Translation: A Case Study on Attributive Clauses(opens in a new tab) (March 2023)

• Knowledge-augmented Frame Semantic Parsing with Hybrid Prompt-tuning(opens in a new tab) (March 2023)

• Debiasing Scores and Prompts of 2D Diffusion for Robust Text-to-3D Generation(opens in a new tab) (March 2023)

• Zero-shot Model Diagnosis(opens in a new tab) (March 2023)

• Prompting Large Language Models to Generate Code-Mixed Texts: The Case of South East Asian Languages(opens in a new tab) (March 2023)

• SPeC: A Soft Prompt-Based Calibration on Mitigating Performance Variability in Clinical Notes Summarization(opens in a new tab) (March 2023)

• Large Language Models and Simple, Stupid Bugs(opens in a new tab) (March 2023)

• Can Generative Pre-trained Transformers (GPT) Pass Assessments in Higher Education Programming Courses?(opens in a new tab) (Mar 2023)

• SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models(opens in a new tab) (Mar 2023)

• ICL-D3IE: In-Context Learning with Diverse Demonstrations Updating for Document Information Extraction(opens in a new tab) (March 2023)

• MathPrompter: Mathematical Reasoning using Large Language Models(opens in a new tab) (March 2023)

• Prompt-Based Learning for Thread Structure Prediction in Cybersecurity Forums(opens in a new tab) (March 2023)

• Choice Over Control: How Users Write with Large Language Models using Diegetic and Non-Diegetic Prompting(opens in a new tab) (March 2023)

• Prompting Large Language Models with Answer Heuristics for Knowledge-based Visual Question Answering(opens in a new tab) (March 2023)

• Soft Prompt Guided Joint Learning for Cross-Domain Sentiment Analysis(opens in a new tab) (March 2023)

• SpeechPrompt v2: Prompt Tuning for Speech Classification Tasks(opens in a new tab) (March 2023)

• Goal Driven Discovery of Distributional Differences via Language Descriptions(opens in a new tab) (Feb 2023)

• Navigating the Grey Area: Expressions of Overconfidence and Uncertainty in Language Models(opens in a new tab) (Feb 2023)

• TabGenie: A Toolkit for Table-to-Text Generation(opens in a new tab) (Feb 2023)

• SGL-PT: A Strong Graph Learner with Graph Prompt Tuning(opens in a new tab) (Feb 2023)

• Few-Shot Table-to-Text Generation with Prompt-based Adapter(opens in a new tab) (Feb 2023)

• Language Models Are Few-shot Learners for Prognostic Prediction(opens in a new tab) (Feb 2023)

• STA: Self-controlled Text Augmentation for Improving Text Classifications(opens in a new tab) (Feb 2023)

• Check Your Facts and Try Again: Improving Large Language Models with External Knowledge and Automated Feedback(opens in a new tab) (Feb 2023)

• How Generative AI models such as ChatGPT can be (Mis)Used in SPC Practice, Education, and Research? An Exploratory Study(opens in a new tab) (Feb 2023)

• Grimm in Wonderland: Prompt Engineering with Midjourney to Illustrate Fairytales(opens in a new tab) (Feb 2023)

• LabelPrompt: Effective Prompt-based Learning for Relation Classification(opens in a new tab) (Feb 2023)

• Language Model Crossover: Variation through Few-Shot Prompting(opens in a new tab) (Feb 2023)

• Prompt Tuning of Deep Neural Networks for Speaker-adaptive Visual Speech Recognition(opens in a new tab) (Feb 2023)

• The Capacity for Moral Self-Correction in Large Language Models(opens in a new tab) (Feb 2023)

• Prompting for Multimodal Hateful Meme Classification(opens in a new tab) (Feb 2023)

• PLACES: Prompting Language Models for Social Conversation Synthesis(opens in a new tab) (Feb 2023)

• Commonsense-Aware Prompting for Controllable Empathetic Dialogue Generation(opens in a new tab) (Feb 2023)

• Crawling the Internal Knowledge-Base of Language Models(opens in a new tab) (Jan 2023)

• Legal Prompt Engineering for Multilingual Legal Judgement Prediction(opens in a new tab) (Dec 2022)

• Investigating Prompt Engineering in Diffusion Models(opens in a new tab) (Nov 2022)

• Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answering(opens in a new tab) (Sep 2022)

• Conversing with Copilot: Exploring Prompt Engineering for Solving CS1 Problems Using Natural Language(opens in a new tab) (Oct 2022)

• Piloting Copilot and Codex: Hot Temperature, Cold Prompts, or Black Magic?(opens in a new tab) (Oct 2022)

• Plot Writing From Scratch Pre-Trained Language Models(opens in a new tab) (July 2022)

• Survey of Hallucination in Natural Language Generation(opens in a new tab) (Feb 2022)

論文匯編

• Chain-of-Thought Papers(opens in a new tab)

• Papers with Code(opens in a new tab)

• Prompt Papers

參考資料

1. [1] Prompt Engineering https://lilianweng.github.io/posts/2023-03-15-prompt-engineering/#chain-of-thought-cot

2. [2] Prompt Engineering Guide https://www.promptingguide.ai/zh

3. [3] https://platform.openai.com/playgroundyground

4. [4] [2201.11903] Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (arxiv.org)

5. [5] [2211.01910] Large Language Models Are Human-Level Prompt Engineers (arxiv.org)

6. [6] Automatic Prompt Augmentation and Selection with Chain-of-Thought from Labeled Data

關(guān)于NLP那些你不知道的事-VIP社群是高質(zhì)量圈子，無(wú)廣告營(yíng)銷及雜亂等無(wú)用內(nèi)容！只有一個(gè)人走得更快，一群人才能走得更遠(yuǎn)，加入VIP社區(qū)，不斷提高學(xué)習(xí)水平。

首先，因?yàn)閂IP社區(qū)（知識(shí)星球）付費(fèi)。所以我們必須對(duì)得起付費(fèi)用戶。努力編寫有價(jià)值的文章，這些文章比公眾號(hào)上的免費(fèi)材料和文件更全面、質(zhì)量更高。

AIGC的發(fā)展每日一個(gè)樣，如果不主動(dòng)學(xué)習(xí)起來(lái)，不用等到35歲，估計(jì)掉隊(duì)只是分分鐘的事情，像ChatGPT、Stable diffusion等新技術(shù)是否了解，AIGC 新理念是否有接觸。

目前AIGC技術(shù)高速迭代，基本每天一個(gè) LLMs 或者新idea，因此需要我們時(shí)刻保持主動(dòng)學(xué)習(xí)，帶有學(xué)習(xí)新技術(shù)新理念的心態(tài)，我們星球也在持續(xù)加碼 AIGC 新理念新技術(shù)分享，像ChatGPT、Stable diffusion 等新理念/新技術(shù) 會(huì)找一些優(yōu)秀的學(xué)習(xí)資料分享在星球！讓星球的小伙伴努力跟上 AIGC 發(fā)展的腳步，不被時(shí)代所拋棄！