View source for Prompt engineering

{{unreliable sources|date=October 2024}}
{{Use mdy dates|date=January 2025}}{{Short description|Structuring text as input to generative artificial intelligence}}
{{Distinguish|text=[[Command prompt]]}}'''Prompt engineering''' is the process of structuring or crafting an instruction in order to produce the best possible output from a [[generative artificial intelligence]] (AI) model.<ref name="Genkina">{{cite web|first=Dina |last=Genkina |title=AI Prompt Engineering is Dead: Long live AI prompt engineering |url=https://spectrum.ieee.org/prompt-engineering-is-dead |website=[[IEEE Spectrum]] |date=2024-03-06 |access-date=2025-01-18}}</ref>

A ''prompt'' is [[natural language]] text describing the task that an AI should perform.<ref name="language-models-are-multitask">{{cite web |last1=Radford |first1=Alec |last2=Wu |first2=Jeffrey |last3=Child |first3=Rewon |last4=Luan |first4=David |last5=Amodei |first5=Dario |author-link5=Dario Amodei |last6=Sutskever |first6=Ilya |author-link6=Ilya Sutskever |year=2019 |title=Language Models are Unsupervised Multitask Learners |url=https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf |publisher=OpenAI |quote="We demonstrate language models can perform down-stream tasks in a zero-shot setting – without any parameter or architecture modification"}}</ref> A prompt for a text-to-text [[Large language model|language model]] can be a query, a command, or a longer statement including context, instructions, and conversation history. Prompt engineering may involve phrasing a query, specifying a style, choice of words and grammar,<ref name="auto1">{{Cite book |last1=Wahle |first1=Jan Philip |title=Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing |last2=Ruas |first2=Terry |last3=Xu |first3=Yang |last4=Gipp |first4=Bela |date=2024 |publisher=Association for Computational Linguistics |editor-last=Al-Onaizan |editor-first=Yaser |location=Miami, Florida, USA |pages=11004–11033 |chapter=Paraphrase Types Elicit Prompt Engineering Capabilities |doi=10.18653/v1/2024.emnlp-main.617 |editor2-last=Bansal |editor2-first=Mohit |editor3-last=Chen |editor3-first=Yun-Nung |chapter-url=https://aclanthology.org/2024.emnlp-main.617/ |arxiv=2406.19898}}</ref> providing relevant context, or describing a character for the AI to mimic.<ref name="Genkina"/>

When communicating with a [[text-to-image]] or a text-to-audio model, a typical prompt is a description of a desired output such as "a high-quality photo of an astronaut riding a horse"<ref>{{Cite web|last=Heaven|first=Will Douglas|title=This horse-riding astronaut is a milestone on AI's long road towards understanding|url=https://www.technologyreview.com/2022/04/06/1049061/dalle-openai-gpt3-ai-agi-multimodal-image-generation/|website=MIT Technology Review|access-date=2023-08-14|date=April 6, 2022}}</ref> or "Lo-fi slow BPM electro chill with organic samples".<ref>{{cite web|last=Wiggers|first=Kyle|title=Meta open sources an AI-powered music generator|url=https://techcrunch.com/2023/06/12/meta-open-sources-an-ai-powered-music-generator/|publisher=TechCrunch|date=2023-06-12|access-date=2023-08-15|quote=Next, I gave a more complicated prompt to attempt to throw MusicGen for a loop: "Lo-fi slow BPM electro chill with organic samples."}}</ref> Prompting a text-to-image model may involve adding, removing, emphasizing, and re-ordering words to achieve a desired subject, style,<ref name="diab">{{Cite web|url=https://cdn.openart.ai/assets/Stable%20Diffusion%20Prompt%20Book%20From%20OpenArt%2011-13.pdf |title=Stable Diffusion Prompt Book |last1=Diab |first1=Mohamad |last2=  Herrera |first2=Julian |last3=Chernow |first3=Bob |date=2022-10-28 |access-date=2023-08-07 |quote="Prompt engineering is the process of structuring words that can be interpreted and understood by a ''text-to-image'' model. Think of it as the language you need to speak in order to tell an AI model what to draw."}}</ref> layout, lighting,<ref>{{Cite web|url=https://claid.ai/blog/article/prompt-guide/|title=How to Write AI Photoshoot Prompts: A Guide for Better Product Photos|website=claid.ai|access-date=June 12, 2023|date=June 12, 2023}}</ref> and aesthetic.

== History ==
In 2018, researchers first proposed that all previously separate tasks in [[natural language processing]] (NLP) could be cast as a question-answering problem over a context. In addition, they trained a first single, joint, multi-task model that would answer any task-related question like "What is the sentiment" or "Translate this sentence to German" or "Who is the president?"<ref>{{cite arXiv | eprint=1806.08730 | last1=McCann | first1=Bryan | last2=Shirish | first2=Nitish | last3=Xiong | first3=Caiming | last4=Socher | first4=Richard | title=The Natural Language Decathlon: Multitask Learning as Question Answering | year=2018 | class=cs.CL }}</ref>

The [[AI boom]] saw an increase in the amount of "prompting technique" to get the model to output the desired outcome and avoid [[Hallucination (artificial intelligence)|nonsensical output]], a process characterized by [[Trial and error|trial-and-error]].<ref>{{Cite journal |last1=Knoth |first1=Nils |last2=Tolzin |first2=Antonia |last3=Janson |first3=Andreas |last4=Leimeister |first4=Jan Marco |date=2024-06-01 |title=AI literacy and its implications for prompt engineering strategies |journal=Computers and Education: Artificial Intelligence |volume=6 |pages=100225 |doi=10.1016/j.caeai.2024.100225 |issn=2666-920X|doi-access=free }}</ref> After the release of [[ChatGPT]] in 2022, prompt engineering was soon seen as an important business skill, albeit one with an uncertain economic future.<ref name="Genkina"/>

A repository for prompts reported that over 2,000 public prompts for around 170 datasets were available in February 2022.<ref>{{cite arXiv | eprint=2202.01279 | last1=Bach | first1=Stephen H. | last2=Sanh | first2=Victor | last3=Yong | first3=Zheng-Xin | last4=Webson | first4=Albert | last5=Raffel | first5=Colin | last6=Nayak | first6=Nihal V. | last7=Sharma | first7=Abheesht | last8=Kim | first8=Taewoon | author9=M Saiful Bari | last10=Fevry | first10=Thibault | last11=Alyafeai | first11=Zaid | last12=Dey | first12=Manan | last13=Santilli | first13=Andrea | last14=Sun | first14=Zhiqing | last15=Ben-David | first15=Srulik | last16=Xu | first16=Canwen | last17=Chhablani | first17=Gunjan | last18=Wang | first18=Han | author19=Jason Alan Fries | last20=Al-shaibani | first20=Maged S. | last21=Sharma | first21=Shanya | last22=Thakker | first22=Urmish | last23=Almubarak | first23=Khalid | last24=Tang | first24=Xiangru | last25=Radev | first25=Dragomir | author26=Mike Tian-Jian Jiang | last27=Rush | first27=Alexander M. | title=PromptSource: An Integrated Development Environment and Repository for Natural Language Prompts | year=2022 | class=cs.LG }}</ref> In 2022, the ''chain-of-thought'' prompting technique was proposed by [[Google]] researchers.<ref name="weipaper">{{cite conference |last1=Wei |first1=Jason |last2=Wang |first2=Xuezhi |last3=Schuurmans |first3=Dale |last4=Bosma |first4=Maarten |last5=Ichter |first5=Brian |last6=Xia |first6=Fei |last7=Chi |first7=Ed H. |last8=Le |first8=Quoc V. |last9=Zhou |first9=Denny |date=31 October 2022 |title=Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |conference=Advances in Neural Information Processing Systems (NeurIPS 2022) |language=en |volume=35 |arxiv=2201.11903}}</ref><ref>{{cite web |last1=Wei |first1=Jason |last2=Zhou |title=Language Models Perform Reasoning via Chain of Thought |url=https://ai.googleblog.com/2022/05/language-models-perform-reasoning-via.html |website=ai.googleblog.com |date=11 May 2022 |access-date=10 March 2023 |language=en}}</ref> In 2023, several text-to-text and text-to-image prompt databases were made publicly available.<ref>{{Cite web|url=https://www.nytimes.com/2023/06/23/technology/ai-chatbot-life-coach.html|title=How to Turn Your Chatbot Into a Life Coach|last=Chen|first=Brian X.|date=2023-06-23|access-date=|website=The New York Times}}</ref><ref>{{Cite news |last=Chen |first=Brian X. |date=2023-05-25 |title=Get the Best From ChatGPT With These Golden Prompts |url=https://www.nytimes.com/2023/05/25/technology/ai-chatbot-chatgpt-prompts.html |url-access=registration |access-date=2023-08-16 |work=The New York Times |language=en-US |issn=0362-4331}}</ref> The Personalized Image-Prompt (PIP) dataset, a generated image-text dataset that has been categorized by 3,115 users, has also been made available publicly in 2024.<ref>{{Cite book |last1=Chen |first1=Zijie |last2=Zhang |first2=Lichao |last3=Weng |first3=Fangsheng |last4=Pan |first4=Lili |last5=Lan |first5=Zhenzhong |chapter=Tailored Visions: Enhancing Text-to-Image Generation with Personalized Prompt Rewriting |date=2024-06-16 |title=2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) |chapter-url=https://doi.org/10.1109/cvpr52733.2024.00738 |publisher=IEEE |pages=7727–7736 |doi=10.1109/cvpr52733.2024.00738|arxiv=2310.08129 |isbn=979-8-3503-5300-6 }}</ref>

== Text-to-text ==
Multiple distinct prompt engineering techniques have been published.

=== Chain-of-thought ===
{{see also|Reflection (artificial intelligence)}}

According to Google Research, ''chain-of-thought'' (CoT) prompting is a technique that allows [[large language models]] (LLMs) to solve a problem as a series of intermediate steps before giving a final answer. In 2022, [[Google Brain]] reported that chain-of-thought prompting improves reasoning ability by inducing the model to answer a multi-step problem with steps of reasoning that mimic a [[train of thought]].<ref name="weipaper"/><ref>{{Cite web |url=https://ai.googleblog.com/2022/04/pathways-language-model-palm-scaling-to.html |date=2022-04-04|title= Pathways Language Model (PaLM): Scaling to 540 Billion Parameters for Breakthrough Performance |author=Sharan Narang and Aakanksha Chowdhery}}</ref> Chain-of-thought techniques were developed to help LLMs handle multi-step reasoning tasks, such as [[arithmetic]] or [[commonsense reasoning]] questions.<ref>{{cite web |last1=Dang |first1=Ekta |title=Harnessing the power of GPT-3 in scientific research |url=https://venturebeat.com/ai/harnessing-the-power-of-gpt-3-in-scientific-research/ |website=VentureBeat |access-date=10 March 2023 |date=8 February 2023}}</ref><ref>{{cite web |last1=Montti |first1=Roger |title=Google's Chain of Thought Prompting Can Boost Today's Best Algorithms |url=https://www.searchenginejournal.com/google-chain-of-thought-prompting/450106/ |website=Search Engine Journal |access-date=10 March 2023 |language=en |date=13 May 2022}}</ref>

For example, given the question, "Q: The cafeteria had 23 apples. If they used 20 to make lunch and bought 6 more, how many apples do they have?", Google claims that a CoT prompt might induce the LLM to answer "A: The cafeteria had 23 apples originally. They used 20 to make lunch. So they had 23 - 20 = 3. They bought 6 more apples, so they have 3 + 6 = 9. The answer is 9."<ref name="weipaper"/> When applied to [[PaLM]], a 540 billion parameter [[language model]], according to Google, CoT prompting significantly aided the model, allowing it to perform comparably with task-specific [[fine-tuning (machine learning)|fine-tuned]] models on several tasks, achieving [[state of the art|state-of-the-art]] results at the time on the GSM8K [[mathematical reasoning]] [[benchmark (computing)|benchmark]].<ref name="weipaper" /> It is possible to fine-tune models on CoT reasoning datasets to enhance this capability further and stimulate better [[interpretability]].<ref>{{cite arXiv |eprint=2210.11416 |class=cs.LG |first1=Hyung Won |last1=Chung |first2=Le |last2=Hou |title=Scaling Instruction-Finetuned Language Models |date=2022 |last3=Longpre |first3=Shayne |last4=Zoph |first4=Barret |last5=Tay |first5=Yi |last6=Fedus |first6=William |last7=Li |first7=Yunxuan |last8=Wang |first8=Xuezhi |last9=Dehghani |first9=Mostafa |last10=Brahma |first10=Siddhartha |last11=Webson |first11=Albert |last12=Gu |first12=Shixiang Shane |last13=Dai |first13=Zhuyun |last14=Suzgun |first14=Mirac |last15=Chen |first15=Xinyun |last16=Chowdhery |first16=Aakanksha |last17=Castro-Ros |first17=Alex |last18=Pellat |first18=Marie |last19=Robinson |first19=Kevin |last20=Valter |first20=Dasha |last21=Narang |first21=Sharan |last22=Mishra |first22=Gaurav |last23=Yu |first23=Adams |last24=Zhao |first24=Vincent |last25=Huang |first25=Yanping |last26=Dai |first26=Andrew |last27=Yu |first27=Hongkun |last28=Petrov |first28=Slav |last29=Chi |first29=Ed H. |last30=Dean |first30=Jeff |last31=Devlin |first31=Jacob |last32=Roberts |first32=Adam |last33=Zhou |first33=Denny |last34=Le |first34=Quoc V. |last35=Wei |first35=Jason}}</ref><ref>{{cite web |last1=Wei |first1=Jason |last2=Tay |first2=Yi |date=29 November 2022 |title=Better Language Models Without Massive Compute |url=https://ai.googleblog.com/2022/11/better-language-models-without-massive.html |access-date=10 March 2023 |website=ai.googleblog.com |language=en}}</ref>

An example of a CoT prompting:<ref name="KojimaStepByStep" />
    Q: {question}
    A: Let's think step by step.

As originally proposed by Google,<ref name="weipaper" /> each CoT prompt included a few Q&A examples. This made it a ''few-shot'' prompting technique. However, according to researchers at Google and the [[University of Tokyo]], simply appending the words "Let's think step-by-step",<ref name="KojimaStepByStep">{{cite arXiv | eprint=2205.11916 | last1=Kojima | first1=Takeshi | author2=Shixiang Shane Gu | last3=Reid | first3=Machel | last4=Matsuo | first4=Yutaka | last5=Iwasawa | first5=Yusuke | title=Large Language Models are Zero-Shot Reasoners | year=2022 | class=cs.CL }}</ref> has also proven effective, which makes CoT a ''zero-shot'' prompting technique. [[OpenAI]] claims that this prompt allows for better scaling as a user no longer needs to formulate many specific CoT Q&A examples.<ref name="venture1">{{cite web |last=Dickson |first=Ben |title=LLMs have not learned our language — we're trying to learn theirs |url=https://venturebeat.com/ai/llms-have-not-learned-our-language-were-trying-to-learn-theirs%EF%BF%BC/ |website=VentureBeat |access-date=10 March 2023 |date=30 August 2022}}</ref>

=== In-context learning ===
''In-context learning'', refers to a model's ability to temporarily learn from prompts. For example, a prompt may include a few examples for a model to learn from, such as asking the model to complete "''<span lang="fr" dir="ltr">maison</span>'' {{arrow}} house, ''<span lang="fr" dir="ltr">chat</span>'' {{arrow}} cat, ''<span lang="fr" dir="ltr">chien</span>'' {{arrow}}" (the expected response being ''dog''),<ref>{{cite arXiv |eprint=2208.01066 |class=cs.CL |first1=Shivam |last1=Garg |first2=Dimitris |last2=Tsipras |title=What Can Transformers Learn In-Context? A Case Study of Simple Function Classes |last3=Liang |first3=Percy |last4=Valiant |first4=Gregory |year=2022}}</ref> an approach called ''few-shot learning''.<ref>{{Cite journal |last1=Brown |first1=Tom |last2=Mann |first2=Benjamin |last3=Ryder |first3=Nick |last4=Subbiah |first4=Melanie |last5=Kaplan |first5=Jared D. |last6=Dhariwal |first6=Prafulla |last7=Neelakantan |first7=Arvind |year=2020 |title=Language models are few-shot learners |journal=Advances in Neural Information Processing Systems |volume=33 |pages=1877–1901 |arxiv=2005.14165}}</ref>

In-context learning is an [[large language model emergent abilities|emergent ability]]<ref name="2022_EmergentAbilities">{{cite arXiv |eprint=2206.07682 |class=cs.CL |first1=Jason |last1=Wei |first2=Yi |last2=Tay |title=Emergent Abilities of Large Language Models |date=31 August 2022 |last3=Bommasani |first3=Rishi |last4=Raffel |first4=Colin |last5=Zoph |first5=Barret |last6=Borgeaud |first6=Sebastian |last7=Yogatama |first7=Dani |last8=Bosma |first8=Maarten |last9=Zhou |first9=Denny |last10=Metzler |first10=Donald |last11=Chi |first11=Ed H. |last12=Hashimoto |first12=Tatsunori |last13=Vinyals |first13=Oriol |last14=Liang |first14=Percy |last15=Dean |first15=Jeff |last16=Fedus |first16=William |quote="In prompting, a pre-trained language model is given a prompt (e.g. a natural language instruction) of a task and completes the response without any further training or gradient updates to its parameters... The ability to perform a task via few-shot prompting is emergent when a model has random performance until a certain scale, after which performance increases to well-above random"}}</ref> of large language models. It is an emergent property of model scale, meaning that [[Broken Neural Scaling Law|breaks]]<ref>{{Citation |last1=Caballero |first1=Ethan |title=Broken Neural Scaling Laws |date=2023-07-24 |arxiv=2210.14891 |last2=Gupta |first2=Kshitij |last3=Rish |first3=Irina |last4=Krueger |first4=David}}</ref> in downstream scaling laws occur, leading to its efficacy increasing at a different rate in larger models than in smaller models.<ref>{{cite arXiv |eprint=2206.07682 |class=cs.CL |first1=Jason |last1=Wei |first2=Yi |last2=Tay |title=Emergent Abilities of Large Language Models |date=31 August 2022 |last3=Bommasani |first3=Rishi |last4=Raffel |first4=Colin |last5=Zoph |first5=Barret |last6=Borgeaud |first6=Sebastian |last7=Yogatama |first7=Dani |last8=Bosma |first8=Maarten |last9=Zhou |first9=Denny |last10=Metzler |first10=Donald |last11=Chi |first11=Ed H. |last12=Hashimoto |first12=Tatsunori |last13=Vinyals |first13=Oriol |last14=Liang |first14=Percy |last15=Dean |first15=Jeff |last16=Fedus |first16=William}}</ref><ref name="weipaper" /> Unlike training and [[Fine-tuning (deep learning)|fine-tuning]], which produce lasting changes, in-context learning is temporary.<ref>{{cite web |last1=Musser |first1=George |author-link=George Musser |title=How AI Knows Things No One Told It |url=https://www.scientificamerican.com/article/how-ai-knows-things-no-one-told-it/ |access-date=17 May 2023 |website=[[Scientific American]] |quote="By the time you type a query into ChatGPT, the network should be fixed; unlike humans, it should not continue to learn. So it came as a surprise that LLMs do, in fact, learn from their users' prompts—an ability known as in-context learning."}}</ref> Training models to perform in-context learning can be viewed as a form of [[Meta-learning (computer science)|meta-learning]], or "learning to learn".<ref>{{Cite journal |date=2022 |title=What Can Transformers Learn In-Context? A Case Study of Simple Function Classes |url= |journal=NeurIPS |arxiv=2208.01066 |quote=Training a model to perform in-context learning can be viewed as an instance of the more general learning-to-learn or meta-learning paradigm |last1=Garg |first1=Shivam |last2=Tsipras |first2=Dimitris |last3=Liang |first3=Percy |last4=Valiant |first4=Gregory }}</ref>

=== Self-consistency decoding ===
''Self-consistency decoding''<ref>{{cite arXiv |last1=Wang |first1=Xuezhi |last2=Wei |first2=Jason |last3=Schuurmans |first3=Dale |last4=Le |first4=Quoc |last5=Chi |first5=Ed |last6=Narang |first6=Sharan |last7=Chowdhery |first7=Aakanksha |last8=Zhou |first8=Denny |date=2022-03-01 |title=Self-Consistency Improves Chain of Thought Reasoning in Language Models |class=cs.CL |eprint=2203.11171}}</ref> performs several chain-of-thought rollouts, then selects the most commonly reached conclusion out of all the rollouts. If the rollouts disagree by a lot, a human can be queried for the correct chain of thought.<ref>{{cite arXiv |last1=Diao |first1=Shizhe |last2=Wang |first2=Pengcheng |last3=Lin |first3=Yong |last4=Zhang |first4=Tong |date=2023-02-01 |title=Active Prompting with Chain-of-Thought for Large Language Models |class=cs.CL |eprint=2302.12246 }}</ref>

=== Tree-of-thought ===
''Tree-of-thought prompting'' generalizes chain-of-thought by prompting the model to generate one or more "possible next steps", and then running the model on each of the possible next steps by [[Breadth-first search|breadth-first]], [[Beam search|beam]], or some other method of tree search.<ref>{{Cite arXiv |last1=Yao |first1=Shunyu |last2=Yu |first2=Dian |last3=Zhao |first3=Jeffrey |last4=Shafran |first4=Izhak |last5=Griffiths |first5=Thomas L. |last6=Cao |first6=Yuan |last7=Narasimhan |first7=Karthik |date=2023-05-17 |title=Tree of Thoughts: Deliberate Problem Solving with Large Language Models |class=cs.CL |eprint=2305.10601 }}</ref> The LLM has additional modules that can converse the history of the problem-solving process to the LLM, which allows the system to 'backtrack steps' the problem-solving process.

=== Prompting to disclose uncertainty ===
By default, the output of language models may not contain estimates of uncertainty. The model may output text that appears confident, though the underlying token predictions have low [[Likelihood function|likelihood]] scores. Large language models like [[GPT-4]] can have accurately [[Calibration (statistics)|calibrated]] likelihood scores in their token predictions,<ref>{{Cite arXiv |eprint=2303.08774 |class=cs.CL |author=((OpenAI and over 200 people))|title=GPT-4 Technical Report |date=2023-03-27}} ''[See Figure 8.]''</ref> and so the model output uncertainty can be directly estimated by reading out the token prediction likelihood scores.

=== Prompting to estimate model sensitivity ===
Research consistently demonstrates that LLMs are highly sensitive to subtle variations in prompt formatting, structure, and linguistic properties. Some studies have shown up to 76 accuracy points across formatting changes in few-shot settings.<ref name="auto">{{cite arXiv |last1=Sclar |first1=Melanie |title=Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formatting |date=2024-07-01 |eprint=2310.11324 |last2=Choi |first2=Yejin |last3=Tsvetkov |first3=Yulia |last4=Suhr |first4=Alane|class=cs.CL }}</ref> Linguistic features significantly influence prompt effectiveness—such as morphology, syntax, and lexico-semantic changes—which meaningfully enhance task performance across a variety of tasks.<ref name="auto1"/><ref>{{Cite journal |last1=Leidinger |first1=Alina |last2=van Rooij |first2=Robert |last3=Shutova |first3=Ekaterina |date=2023 |editor-last=Bouamor |editor-first=Houda |editor2-last=Pino |editor2-first=Juan |editor3-last=Bali |editor3-first=Kalika |title=The language of prompting: What linguistic properties make a prompt successful? |url=https://aclanthology.org/2023.findings-emnlp.618/ |journal=Findings of the Association for Computational Linguistics: EMNLP 2023 |location=Singapore |publisher=Association for Computational Linguistics |pages=9210–9232 |doi=10.18653/v1/2023.findings-emnlp.618|arxiv=2311.01967 }}</ref> Clausal syntax, for example, improves consistency and reduces uncertainty in knowledge retrieval.<ref>{{Cite book |last1=Linzbach |first1=Stephan |last2=Dimitrov |first2=Dimitar |last3=Kallmeyer |first3=Laura |last4=Evang |first4=Kilian |last5=Jabeen |first5=Hajira |last6=Dietze |first6=Stefan |chapter=Dissecting Paraphrases: The Impact of Prompt Syntax and supplementary Information on Knowledge Retrieval from Pretrained Language Models |date=June 2024 |editor-last=Duh |editor-first=Kevin |editor2-last=Gomez |editor2-first=Helena |editor3-last=Bethard |editor3-first=Steven |title=Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) |chapter-url=https://aclanthology.org/2024.naacl-long.201/ |location=Mexico City, Mexico |publisher=Association for Computational Linguistics |pages=3645–3655 |doi=10.18653/v1/2024.naacl-long.201|arxiv=2404.01992 }}</ref> This sensitivity persists even with larger model sizes, additional few-shot examples, or instruction tuning.

To address sensitivity of models and make them more robust, several methods have been proposed. FormatSpread facilitates systematic analysis by evaluating a range of plausible prompt formats, offering a more comprehensive performance interval.<ref name="auto"/> Similarly, PromptEval estimates performance distributions across diverse prompts, enabling robust metrics such as performance quantiles and accurate evaluations under constrained budgets.<ref>{{cite arXiv|last1=Polo |first1=Felipe Maia |title=Efficient multi-prompt evaluation of LLMs |date=2024-10-30 |eprint=2405.17202 |last2=Xu |first2=Ronald |last3=Weber |first3=Lucas |last4=Silva |first4=Mírian |last5=Bhardwaj |first5=Onkar |last6=Choshen |first6=Leshem |last7=de Oliveira |first7=Allysson Flavio Melo |last8=Sun |first8=Yuekai |last9=Yurochkin |first9=Mikhail|class=cs.CL }}</ref>

=== Automatic prompt generation ===

==== Retrieval-augmented generation ====
{{Main article|Retrieval-augmented generation}}
[[File:RAG schema.svg|thumb|Two-phase process of document retrieval using dense [[Word embedding|embeddings]] and LLM for answer formulation]]
Retrieval-augmented generation (RAG) is a two-phase process involving [[document retrieval]] and answer generation by a large language model. The initial phase uses dense embeddings to retrieve documents. This retrieval can be based on a variety of database formats depending on the use case, such as a [[vector database]], summary index, tree index, or keyword table index.<ref>{{Cite web |title=How Each Index Works - LlamaIndex 🦙 v0.10.17 |url=https://docs.llamaindex.ai/en/v0.10.17/module_guides/indexing/index_guide.html |access-date=2024-04-08 |website=docs.llamaindex.ai}}</ref> In response to a query, a document retriever selects the most relevant documents. This relevance is typically determined by first encoding both the query and the documents into vectors, then identifying documents whose vectors are closest in [[Euclidean distance]] to the query vector. Following document retrieval, the LLM generates an output that incorporates information from both the query and the retrieved documents.<ref>{{Cite journal |last1=Lewis |first1=Patrick |last2=Perez |first2=Ethan |last3=Piktus |first3=Aleksandra |last4=Petroni |first4=Fabio |last5=Karpukhin |first5=Vladimir |last6=Goyal |first6=Naman |last7=Küttler |first7=Heinrich |last8=Lewis |first8=Mike |last9=Yih |first9=Wen-tau |last10=Rocktäschel |first10=Tim |last11=Riedel |first11=Sebastian |last12=Kiela |first12=Douwe |date=2020 |title=Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks |url=https://proceedings.neurips.cc/paper/2020/hash/6b493230205f780e1bc26945df7481e5-Abstract.html |journal=Advances in Neural Information Processing Systems |publisher=Curran Associates, Inc. |volume=33 |pages=9459–9474 |arxiv=2005.11401}}</ref> RAG can also be used as a few-shot learner.

==== Graph retrieval-augmented generation ====

[[File:GraphRAG.svg|thumb|GraphRAG with a knowledge graph combining access patterns for unstructured, structured, and mixed data]]

GraphRAG<ref>{{citation |last1=Larson |first1=Jonathan |title=GraphRAG: Unlocking LLM discovery on narrative private data |date=February 13, 2024 |url=https://www.microsoft.com/en-us/research/blog/graphrag-unlocking-llm-discovery-on-narrative-private-data/ |publisher=Microsoft |last2=Truitt |first2=Steven}}</ref> (coined by [[Microsoft Research]]) is a technique that extends RAG with the use of a knowledge graph (usually, LLM-generated) to allow the model to connect disparate pieces of information, synthesize insights, and holistically understand summarized semantic concepts over large data collections. It was shown to be effective on datasets like the Violent Incident Information from News Articles (VIINA).<ref>{{cite arXiv | date=2024 |title=From Local to Global: A Graph RAG Approach to Query-Focused Summarization |eprint=2404.16130 |last1=Edge |first1=Darren |last2=Trinh |first2=Ha |last3=Cheng |first3=Newman |last4=Bradley |first4=Joshua |last5=Chao |first5=Alex |last6=Mody |first6=Apurva |last7=Truitt |first7=Steven |last8=Larson |first8=Jonathan |class=cs.CL }}</ref>

Earlier work showed the effectiveness of using a [[knowledge graph]] for question answering using text-to-query generation.<ref>{{cite arXiv | date=2023 | title=A Benchmark to Understand the Role of Knowledge Graphs on Large Language Model's Accuracy for Question Answering on Enterprise SQL Databases |eprint=2311.07509 | last1=Sequeda | first1=Juan | last2=Allemang | first2=Dean | last3=Jacob | first3=Bryon | class=cs.AI }}</ref> These techniques can be combined to search across both unstructured and structured data, providing expanded context, and improved ranking.

==== Using language models to generate prompts ====
Large language models (LLM) themselves can be used to compose prompts for large language models.<ref>{{cite arXiv |last1=Singh |first1=Chandan |last2=Morris |first2=John |last3=Aneja |first3=Jyoti |last4=Rush |first4=Alexander |last5=Gao |first5=Jianfeng |title=Explaining Patterns in Data with Language Models via Interpretable Autoprompting |date=October 4, 2022 |class=cs.LG |eprint=2210.01848 }}</ref> The ''automatic prompt engineer'' algorithm uses one LLM to [[beam search]] over prompts for another LLM:<ref>{{cite arXiv |last1=Zhou |first1=Yongchao |last2=Ioan Muresanu |first2=Andrei |last3=Han |first3=Ziwen |last4=Paster |first4=Keiran |last5=Pitis |first5=Silviu |last6=Chan |first6=Harris |last7=Ba |first7=Jimmy |date=2022-11-01 |title=Large Language Models Are Human-Level Prompt Engineers |class=cs.LG |eprint=2211.01910}}</ref><ref>{{Cite journal |last1=Pryzant |first1=Reid |last2=Iter |first2=Dan |last3=Li |first3=Jerry |last4=Lee |first4=Yin Tat |last5=Zhu |first5=Chenguang |last6=Zeng |first6=Michael |date=2023 |title=Automatic Prompt Optimization with "Gradient Descent" and Beam Search |journal=Microsoft Azure AI |arxiv=2305.03495}}</ref>

* There are two LLMs. One is the target LLM, and another is the prompting LLM.
* Prompting LLM is presented with example input-output pairs, and asked to generate instructions that could have caused a model following the instructions to generate the outputs, given the inputs. 
* Each of the generated instructions is used to prompt the target LLM, followed by each of the inputs. The log-probabilities of the outputs are computed and added. This is the score of the instruction.
* The highest-scored instructions are given to the prompting LLM for further variations.
* Repeat until some stopping criteria is reached, then output the highest-scored instructions.
CoT examples can be generated by LLM themselves. In "auto-CoT",<ref>{{cite arXiv |last1=Zhang |first1=Zhuosheng |last2=Zhang |first2=Aston |last3=Li |first3=Mu |last4=Smola |first4=Alex |date=2022-10-01 |title=Automatic Chain of Thought Prompting in Large Language Models |class=cs.CL |eprint=2210.03493}}</ref> a library of questions are converted to vectors by a model such as [[BERT (language model)|BERT]]. The question vectors are [[Cluster analysis|clustered]]. Questions nearest to the centroids of each cluster are selected. An LLM does zero-shot CoT on each question. The resulting CoT examples are added to the dataset. When prompted with a new question, CoT examples to the nearest questions can be retrieved and added to the prompt.

== Text-to-image ==
{{see also|Artificial intelligence art#Prompt engineering and sharing}}
[[File:Fooocus 2.5.5 screenshot showing the prompt section.webp|thumb|upright=1.2|Example of prompt engineering for text-to-image generation, with [[Fooocus]]]]
In 2022, [[text-to-image]] models like [[DALL-E 2]], [[Stable Diffusion]], and [[Midjourney]] were released to the public.<ref>{{Cite web |last=Monge |first=Jim Clyde |date=2022-08-25 |title=Dall-E2 VS Stable Diffusion: Same Prompt, Different Results |url=https://medium.com/mlearning-ai/dall-e2-vs-stable-diffusion-same-prompt-different-results-e795c84adc56 |access-date=2022-08-31 |website=MLearning.ai |language=en}}</ref> These models take text prompts as input and use them to generate [[Artificial intelligence art|AI-generated images]]. Text-to-image models typically do not understand grammar and sentence structure in the same way as [[large language models]],<ref name="Prompts">{{Cite web |title=Prompts |url=https://docs.midjourney.com/docs/prompts |access-date=2023-08-14 |website=docs.midjourney.com}}</ref> thus may require a different set of prompting techniques.

Text-to-image models do not natively understand negation. The prompt "a party with no cake" is likely to produce an image including a cake.<ref name="Prompts" /> As an alternative, ''negative prompts'' allow a user to indicate, in a separate prompt, which terms should ''not'' appear in the resulting image.<ref>{{Cite web |author=Max Woolf |date=2022-11-28 |title=Stable Diffusion 2.0 and the Importance of Negative Prompts for Good Results |url=https://minimaxir.com/2022/11/stable-diffusion-negative-prompt/ |access-date=2023-08-14}}</ref> Techniques such as framing the normal prompt into a [[Seq2seq|sequence-to-sequence]] language modeling problem can be used to automatically generate an output for the negative prompt.<ref>{{Cite journal |last1=Goldblum |first1=R. |last2=Pillarisetty |first2=R. |last3=Dauphinee |first3=M. J. |last4=Talal |first4=N. |date=1975 |title=Acceleration of autoimmunity in NZB/NZW F1 mice by graft-versus-host disease |journal=Clinical and Experimental Immunology |volume=19 |issue=2 |pages=377–385 |issn=0009-9104 |pmc=1538084 |pmid=2403}}</ref>

{{multiple image
| direction         = vertical
| align             = right
| total_width       = 200
| image1            = Algorithmically-generated landscape artwork of forest with Shinto shrine.png
| image2            = Algorithmically-generated landscape artwork of forest with Shinto shrine using negative prompt for green trees.png
| image3            = Algorithmically-generated landscape artwork of forest with Shinto shrine using negative prompt for round stones.png
| footer            = Demonstration of the effect of negative prompts on images generated with [[Stable Diffusion]]
{{bulleted list|'''Top''': no negative prompt
|'''Centre''': "green trees"
|'''Bottom''': "round stones, round rocks"}}
}}

=== Prompt formats ===

A text-to-image prompt commonly includes a description of the subject of the art, the desired medium (such as ''digital painting'' or ''photography''), style (such as ''hyperrealistic'' or ''pop-art''), lighting (such as ''rim lighting'' or ''crepuscular rays''), color, and texture.<ref>{{Cite web|url=https://stable-diffusion-art.com/prompt-guide/|title=Stable Diffusion prompt: a definitive guide|date=2023-05-14|access-date=2023-08-14}}</ref> Word order also affects the output of a text-to-image prompt. Words closer to the start of a prompt may be emphasized more heavily.<ref name="diab" />

The [[Midjourney]] documentation encourages short, descriptive prompts: instead of "Show me a picture of lots of blooming California poppies, make them bright, vibrant orange, and draw them in an illustrated style with colored pencils", an effective prompt might be "Bright orange California poppies drawn with colored pencils".<ref name="Prompts"/>

=== Artist styles ===

Some text-to-image models are capable of imitating the style of particular artists by name. For example, the phrase ''in the style of Greg Rutkowski'' has been used in Stable Diffusion and Midjourney prompts to generate images in the distinctive style of Polish digital artist [[Greg Rutkowski]].<ref>{{Cite web|url=https://www.technologyreview.com/2022/09/16/1059598/this-artist-is-dominating-ai-generated-art-and-hes-not-happy-about-it/|title=This Artist Is Dominating AI-Generated Art and He's Not Happy About It|last=Heikkilä|first=Melissa|date=2022-09-16|website=MIT Technology Review|access-date=2023-08-14}}</ref> Famous artists such as [[Vincent van Gogh]] and [[Salvador Dalí]] have also been used for styling and testing.<ref>{{Cite web |last=Solomon |first=Tessa |date=2024-08-28 |title=The AI-Powered Ask Dalí and Hello Vincent Installations Raise Uncomfortable Questions about Ventriloquizing the Dead |url=https://www.artnews.com/art-news/issue/salvador-dali-vincent-van-gogh-ai-installations-ethics-1234714954/ |access-date=2025-01-10 |website=ARTnews.com |language=en-US}}</ref>

== Non-text prompts ==

Some approaches augment or replace natural language text prompts with non-text input.

=== Textual inversion and embeddings ===
For text-to-image models, ''textual inversion''<ref>{{cite arXiv |last1=Gal |first1=Rinon |last2=Alaluf |first2=Yuval |last3=Atzmon |first3=Yuval |last4=Patashnik |first4=Or |last5=Bermano |first5=Amit H. |last6=Chechik |first6=Gal |last7=Cohen-Or |first7=Daniel |title=An Image is Worth One Word: Personalizing Text-to-Image Generation using Textual Inversion |year=2022 |class=cs.CV |eprint=2208.01618|quote="Using only 3-5 images of a user-provided concept, like an object or a style, we learn to represent it through new "words" in the embedding space of a frozen text-to-image model."}}</ref> performs an optimization process to create a new [[word embedding]] based on a set of example images. This embedding vector acts as a "pseudo-word" which can be included in a prompt to express the content or style of the examples.

=== Image prompting ===
In 2023, [[Meta Platforms|Meta]]'s AI research released Segment Anything, a [[computer vision]] model that can perform [[image segmentation]] by prompting. As an alternative to text prompts, Segment Anything can accept bounding boxes, segmentation masks, and foreground/background points.<ref name="Kirillov">{{cite arXiv |last1=Kirillov |first1=Alexander |last2=Mintun |first2=Eric |last3=Ravi |first3=Nikhila |last4=Mao |first4=Hanzi |last5=Rolland |first5=Chloe |last6=Gustafson |first6=Laura |last7=Xiao |first7=Tete |last8=Whitehead |first8=Spencer |last9=Berg |first9=Alexander C. |last10=Lo |first10=Wan-Yen |last11=Dollár |first11=Piotr |last12=Girshick |first12=Ross |date=2023-04-01 |title=Segment Anything |class=cs.CV |eprint=2304.02643}}</ref>

=== Using gradient descent to search for prompts ===
In "prefix-tuning",<ref>{{cite book | doi=10.18653/V1/2021.ACL-LONG.353 | chapter=Prefix-Tuning: Optimizing Continuous Prompts for Generation | title=Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers) | year=2021 | last1=Li | first1=Xiang Lisa | last2=Liang | first2=Percy | pages=4582–4597 | s2cid=230433941|quote="In this paper, we propose prefix-tuning, a lightweight alternative to fine-tuning... Prefix-tuning draws inspiration from prompting"}}</ref> "prompt tuning", or "soft prompting",<ref>{{cite book | doi=10.18653/V1/2021.EMNLP-MAIN.243 | chapter=The Power of Scale for Parameter-Efficient Prompt Tuning | title=Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing | year=2021 | last1=Lester | first1=Brian | last2=Al-Rfou | first2=Rami | last3=Constant | first3=Noah | pages=3045–3059 | s2cid=233296808 |arxiv=2104.08691|quote="In this work, we explore "prompt tuning," a simple yet effective mechanism for learning "soft prompts"...Unlike the discrete text prompts used by GPT-3, soft prompts are learned through back-propagation"}}</ref> floating-point-valued vectors are searched directly by [[gradient descent]] to maximize the log-likelihood on outputs.

Formally, let <math>\mathbf{E} = \{\mathbf{e_1}, \dots, \mathbf{e_k}\}</math> be a set of soft prompt tokens (tunable embeddings), while <math>\mathbf{X} = \{\mathbf{x_1}, \dots, \mathbf{x_m}\}</math> and <math>\mathbf{Y} = \{\mathbf{y_1}, \dots, \mathbf{y_n}\}</math> be the token embeddings of the input and output respectively. During training, the tunable embeddings, input, and output tokens are concatenated into a single sequence <math>\text{concat}(\mathbf{E};\mathbf{X};\mathbf{Y})</math>, and fed to the LLMs. The [[Loss function|losses]] are computed over the <math>\mathbf{Y}</math> tokens; the gradients are [[Backpropagation|backpropagated]] to prompt-specific parameters: in prefix-tuning, they are parameters associated with the prompt tokens at each layer; in prompt tuning, they are merely the soft tokens added to the vocabulary.<ref>{{Cite arXiv |title=How Does In-Context Learning Help Prompt Tuning?|eprint= 2302.11521|last1= Sun|first1= Simeng|last2= Liu|first2= Yang|last3= Iter|first3= Dan|last4= Zhu|first4= Chenguang|last5= Iyyer|first5= Mohit|year= 2023|class= cs.CL}}</ref>

More formally, this is prompt tuning. Let an LLM be written as <math>LLM(X) = F(E(X)) </math>, where <math>X</math> is a sequence of linguistic tokens, <math>E</math> is the token-to-vector function, and <math>F</math> is the rest of the model. In prefix-tuning, one provides a set of input-output pairs <math>\{(X^i, Y^i)\}_i</math>, and then use gradient descent to search for <math>\arg\max_{\tilde Z} \sum_i \log Pr[Y^i | \tilde Z \ast E(X^i)]</math>. In words, <math>\log Pr[Y^i | \tilde Z \ast E(X^i)]</math> is the log-likelihood of outputting <math>Y^i</math>, if the model first encodes the input <math>X^i</math> into the vector <math>E(X^i)</math>, then prepend the vector with the "prefix vector" <math>\tilde Z</math>, then apply <math>F</math>.

For prefix tuning, it is similar, but the "prefix vector" <math>\tilde Z</math> is pre-appended to the hidden states in every layer of the model.

An earlier result<ref>{{Cite book |last1=Shin |first1=Taylor |last2=Razeghi |first2=Yasaman |last3=Logan IV |first3=Robert L. |last4=Wallace |first4=Eric |last5=Singh |first5=Sameer |title=Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) |chapter=AutoPrompt: Eliciting Knowledge from Language Models with Automatically Generated Prompts |date=November 2020 |chapter-url=https://aclanthology.org/2020.emnlp-main.346 |location=Online |publisher=Association for Computational Linguistics |pages=4222–4235 |doi=10.18653/v1/2020.emnlp-main.346|s2cid=226222232 |doi-access=free }}</ref> uses the same idea of gradient descent search, but is designed for masked language models like BERT, and searches only over token sequences, rather than numerical vectors. Formally, it searches for <math>\arg\max_{\tilde X} \sum_i \log Pr[Y^i | \tilde X \ast X^i]</math> where <math>\tilde X</math> is ranges over token sequences of a specified length.

== Prompt injection ==
{{Main article|Prompt injection}}
{{see also|SQL injection|Cross-site scripting}}

''Prompt injection'' is a family of related [[computer security exploit]]s carried out by getting a [[machine learning]] model (such as an LLM) which was trained to follow human-given instructions to follow instructions provided by a malicious user. This stands in contrast to the intended operation of instruction-following systems, wherein the ML model is intended only to follow trusted instructions (prompts) provided by the ML model's operator.<ref>{{Cite web |last=Willison |first=Simon |author-link=Simon Willison |date=12 September 2022 |title=Prompt injection attacks against GPT-3 |url=http://simonwillison.net/2022/Sep/12/prompt-injection/ |access-date=2023-02-09 |website=simonwillison.net |language=en-gb}}</ref><ref>{{Cite web |last=Papp |first=Donald |date=2022-09-17 |title=What's Old Is New Again: GPT-3 Prompt Injection Attack Affects AI |url=https://hackaday.com/2022/09/16/whats-old-is-new-again-gpt-3-prompt-injection-attack-affects-ai/ |access-date=2023-02-09 |website=[[Hackaday]] |language=en-US}}</ref><ref>{{Cite web |last=Vigliarolo |first=Brandon |date=19 September 2022 |title=GPT-3 'prompt injection' attack causes bot bad manners |url=https://www.theregister.com/2022/09/19/in_brief_security/ |access-date=2023-02-09 |website=[[The Register]] |language=en}}</ref>

== See also ==
*[[Social engineering (security)]]

== References ==
<references />
{{Scholia|topic}}
{{Generative AI}}
{{Artificial intelligence navbox}}

[[Category:Deep learning]]
[[Category:Machine learning]]
[[Category:Natural language processing]]
[[Category:Unsupervised learning]]
[[Category:2022 neologisms]]
[[Category:Linguistics]]
[[Category:Generative artificial intelligence]]