Prompt Engineering in Large Language

Models

Ggaliwango Marvin , Nakayiza Hellen , Daudi Jjingo ,

and Joyce Nakatumba-Nabende

1 Introduction

1.1 Overview of Large Language Models (LLMs)

A Large Language Model (LLM) is an AI algorithm that employs deep learning

and vast datasets to comprehend, summarize, generate, and predict new content.
LLMs are trained on extensive amounts of text using self-supervised learning and
excel at a wide range of tasks [1]. LLMs are constructed using neural networks
and multiple parameters usually billions of weights and more. These models are
pre-trained on vast amounts of data to help them understand the complexity and
relationships within language. By using techniques such as fine-tuning, in-context
learning, and zero-/one-/few-shot learning, these models can be tailored for specific
tasks [2]. A Large Language Model (LLM) is essentially a transformer-based neural
network that predicts the text that is likely to come next, therefore its sophistication
and performance can be evaluated by the number of parameters it possesses. The
model’s parameters are the factors it takes into account when generating output [3].
LLMs have generated much excitement recently due to their impressive capabilities
as general-purpose computers when conditioned on natural language instructions.
However, the effectiveness of the model is heavily influenced by the quality of the
prompt used to guide it, and most successful prompts have been created manually
by humans [4]. This has led to the rise of prompt engineering as an important skill
set for NLP and AI engineers, conversational AI researchers, and most importantly,

G. Marvin (B) · D. Jjingo · J. Nakatumba-Nabende

Makerere University, Kampala, Uganda
e-mail: [Link]@[Link]
N. Hellen
Muni University, Muni, Uganda

© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024 387
I. J. Jacob et al. (eds.), Data Intelligence and Cognitive Informatics, Algorithms for
Intelligent Systems, [Link]
388 G. Marvin et al.

normal information seekers in various domains like education and health care were
improved efficiency in the use of LLMs which is more valued.

1.2 Importance of Prompt Engineering for LLMs

Prompt engineering is the process of designing and refining input queries, or

“prompts,” to elicit desired responses from Large Language Models (LLMs). Prompts
are crucial in guiding LLMs to generate useful and relevant outputs [5]. By under-
standing how to create effective prompts, information seekers and developers can
improve LLM performance, explore new applications, and save valuable time and
resources. This can be achieved by leveraging good quality prompts to guide the
model toward generating more accurate and useful outputs. This in turn allows infor-
mation seekers and developers to unlock the full potential of LLMs. Prompt engi-
neering permits exploration of new applications for LLMs, saves valuable time and
resources, and provides a form of programming that can customize the outputs and
interactions with an LLM [6]. It offers reusable solutions to common problems faced
in output generation and interaction when working with LLMs. By facilitating the
development of more accurate, context-specific, and nuanced responses from LLMs,
prompt engineering supports the advancement of research and discoveries in the field
of AI most especially conversational AI.
Prompt engineering provides a framework for documenting patterns for struc-
turing prompts to solve a range of problems and allows information seekers and
developers to adapt prompts to different domains. It enables the combination of
multiple prompt patterns to improve LLM outputs and facilitates the transfer of
knowledge between information seekers and developers working with LLMs [7].
Prompt engineering supports the development of more sophisticated and effective
LLM applications, helps developers to better understand the behavior and capabili-
ties of LLMs, and enables information seekers and developers to steer LLMs toward
truthfulness and/or informativeness. It also improves few-shot learning performance
by prepending optimized prompts to standard in-context learning prompts and facil-
itates the development of more effective chatbots, virtual assistants, domain-specific
prompt engineering tools, and other conversational AI systems [8, 9]. Therefore,
prompt engineering supports the advancement of natural language processing (NLP)
tasks by improving LLM performance.
Prompt engineering is likely to play an increasingly important role in unlocking
the full potential of Large Language Models (LLMs) in the near future. If information
seekers and developers and empowered to generate specific language output quickly
and accurately, prompt engineering will become marketable element for increasing
efficiency and streamlining business operations across various domains. This rapidly
growing field may also lead to new job opportunities for those skilled in prompt engi-
neering. As information seekers and developers continue to experiment with more
sophisticated prompts [10, 11], we are likely to see the development of more effi-
cient and understandable user interfaces for controlling LLM outputs. This enhanced
Prompt Engineering in Large Language Models 389

control over LLM outputs will enable developers to fine-tune the generated content
and unlock new applications for LLMs that were previously impossible.

1.3 Research Objective and Motivation

The problem that this chapter addresses is the need for effective prompt engineering
to unlock the full potential of Large Language Models (LLMs). LLMs have gener-
ated much hype in the recent months due to their impressive capabilities as general-
purpose computers when conditioned on natural language instructions [12]. However,
the effectiveness of the model is heavily influenced by the quality of the prompt used
to guide it, and most successful prompts have been created manually by humans.
This has led to the rise of prompt engineering as an important skill set for infor-
mation seekers, AI engineers and researchers to improve and efficiently use LLMs.
Therefore, this study provides understanding of prompt engineering, presents an
overview of the latest prompting techniques, and provides demonstrations and exer-
cises to practice different prompting techniques. It also discusses the current and
future trends of research on LLMs and prompt engineering, including the rise of
automatic instruction generation and selection methods for prompt engineering. By
understanding how to create effective prompts, information seekers and developers
can improve LLM performance, explore new applications, and save valuable time
and resources.

2 Prompt Engineering

What are Prompts?

A prompt is a text-based input that is fed to a language model to guide its output.
A prompt can be audio but, in this case, the audio input would be transcribed into text
and fed to the language model as a text-based prompt. The language model would then
process the text-based prompt and generate an output based on the instructions and
context provided in the prompt [13]. The primary purpose of a prompt is to provide
the language model with instructions and context for achieving a desired task. Prompt
engineering is the practice of developing and optimizing prompts to efficiently use
language models (LMs) for a variety of applications. A prompt is composed of an
instruction and can have input data, context, and some output indicator (Table 1).
390 G. Marvin et al.

Table 1 Examples of prompts for large language models (LLMs)

Prompt Instruction Input Context Output indicator
data
Paint a tall African man Paint NA Tall African NA
dancing to cool Ugandan man dancing to
folk songs cool Ugandan
folk songs
“Write a science fiction Write a science Babies Set in a future Explore the ethical
short story set in a future fiction short story where humans implications of
where humans particularly have colonized terraforming and
babies have colonized Mars potential conflicts
Mars, exploring the ethical between Earth and
implications of Martian societies
terraforming and the
potential conflicts between
Earth and Martian
societies.” (Creative
writing)
“Write a short paragraph Write a short NA Explaining a Provide an example
explaining the Pythagorean paragraph mathematical
theorem and providing an concept
example to demonstrate its
use.” (Mathematics)
“Write a philosophical Write a NA Explore the Incorporate
essay exploring the philosophical concept of free arguments from
concept of free will, essay will determinism,
incorporating arguments compatibilism, and
from determinism, libertarianism
compatibilism, and
libertarianism.”
(Philosophy)
“Generate a technical Generate a NA Feasibility of Include an analysis
report on the feasibility of technical report using nuclear of current
using nuclear fusion as a fusion as a technological
sustainable energy source, sustainable limitations and
including an analysis of energy source potential solutions
current technological
limitations and potential
solutions.” (Engineering)
“Write a historical fiction Write a historical NA Set during the Incorporate
novel set during the French fiction novel French accurate historical
Revolution, incorporating revolution details and explore
accurate historical details themes such as
and exploring themes such social inequality,
as social inequality, political upheaval,
political upheaval, and and personal
personal agency.” agency
(Creative writing)
(continued)
Prompt Engineering in Large Language Models 391

Table 1 (continued)
Prompt Instruction Input Context Output indicator
data
“Generate a Generate a NA Current state of Include an analysis
comprehensive review of comprehensive research on of ethical
the current state of review artificial considerations and
research on artificial intelligence and potential risks
intelligence and its its potential
potential impact on society, impact on
including an analysis of society
ethical considerations and
potential risks.”
(Academia)
“Write a legal brief arguing Write a legal brief NA Argue for the Incorporate
for the protection of protection of relevant case law
privacy rights in the digital privacy rights in and constitutional
age, incorporating relevant the digital age principles
case law and constitutional
principles.” (Law)
“Generate a detailed Generate a NA Environmental Include effects on
analysis of the detailed analysis impact of biodiversity, carbon
environmental impact of large-scale sequestration, and
large-scale deforestation, deforestation soil erosion
including effects on
biodiversity, carbon
sequestration, and soil
erosion.” (Environmental
science)
“Write a screenplay for a Write a NA Set in New York Incorporate
romantic comedy set in screenplay for a City elements of social
New York City, romantic comedy commentary and
incorporating elements of explore themes
social commentary and such as identity and
exploring themes such as self-discovery
identity and
self-discovery.”
(Entertainment)
“Generate a Generate a NA For someone Incorporating both
comprehensive treatment comprehensive with chronic traditional and
plan for someone with treatment plan pain alternative
chronic pain, incorporating therapies and
both traditional and taking into account
alternative therapies and factors such as
taking into account factors comorbid
such as comorbid conditions and
conditions and medication medication
interactions.” (Medicine) interactions
392 G. Marvin et al.

2.1 The Process of Prompt Engineering

Steps involved in creating effective prompts:

1. Defining the goal: The first step in creating an effective prompt for a language
model is to clearly define the goal of the prompt. What do you want the model to
generate as a response to the prompt? Having a clear goal in mind enables you
to focus your efforts and create a more effective prompt [14]. This is important
because it provides direction and purpose for the rest of the prompt engineering
process.
2. Understanding the model’s capabilities: For every language model selected for
prompting, it is very important to understand its limitations and abilities. An
example is clarity on what kind of responses the model generates (text, audio,
images, etc.) The strengths and weaknesses of a given model guide the synthesis
of appropriate prompts that are effectively compatible with the model’s capa-
bilities and bypassing such limitations [15]. Fortunately, LLMs today have been
built to optimize their performance by making use of some online and openly
available tools via relevant APIs, e.g., access to diagraming tools.
3. Choosing the right prompt format: There is a huge impact on the quality of
LLM generated responses based on prompt formats. Clear, precise, and concise
prompting formats empower the model with all necessary details for generating
coherent responses [16]. Selection of appropriate prompt formats enhance natural
language understanding for LLMs hence improving quality of response given.
4. Providing Context: This is one of the most under estimated stages yet most
influential in terms of information accuracy generated by the models. Without
context LLMs provide generic coherent and relevant responses [17]. But with
limited contexts, LLMs provide misinformation for less represented context
within existing data. Context may include additional relevant/related information
about the topic, setting or characters for inference within the prompt. Context
provides deeper metrics for the LLM to understand the prompt and desired output.
5. Testing and Refining: For every synthesized prompt, it is vital to test it with
appropriate LLMs based on the defined goal [18]. The generated LLM response
can then be used to evaluate and refine the prompt before its output in a produc-
tion setting. The whole aim of this testing is to obtain feedback for objective
refining before making any data-driven decision about prompt optimization/
improvement.
While the above procedure gives a structured approach to creating optimal and
efficient prompts for LLMs like ChatGPT, information seekers and developers may
choose to deviate from some of the above steps. This may usually happen if one
has the expertise and extensive experience with a particular prompt engineering
tool or LLM. In this case, the information seeker or developer can majorly rely
on their creativity and intuition to synthesize prompts. The divergence may also
happen cases of LLM testing and highly experimental circumstances. In cases of
design and innovative thinking, validation and verification, information seekers may
Prompt Engineering in Large Language Models 393

take exploratory approaches to prompt engineering. Please note that the synthesized
procedure /framework above is not a set of rules and restrictions and not mandatory to
be followed religiously. However, it is a guide that can help prompt engineers create
effective prompts. Ultimately, the most effective approach to prompt engineering will
depend on the specific goals, audience, and language model being used. A skilled
prompt engineer will use their judgment and expertise to determine the best approach
for each situation.

3 Prompt Engineering Techniques

3.1 Techniques for Optimizing Prompts

Optimizing prompts for language models like ChatGPT involves a combination of

creativity, intuition, and data-driven approaches. By using the techniques described
below, you can create effective and engaging prompts that help users achieve their
goals;
Clearly define the goal of the prompt, understand the capabilities and limitations of
the language model, choose a format that is clear and concise, provide context to help
the model generate more relevant responses, test and refine your prompts to improve
their effectiveness, use engaging language and visuals to capture the user’s attention,
Tailor your prompts to the interests and motivations of your audience, experiment
with different formats and styles to see what works best, gather feedback from users
to improve your prompts, use data-driven approaches to refine and optimize your
prompts, keep your prompts focused and on-topic, avoid using overly complex or
technical language, use examples and analogies to help explain complex concepts,
break down complex topics into smaller, more manageable chunks, use humor and
creativity to make your prompts more engaging, leverage the power of storytelling to
make your prompts more memorable, use repetition and reinforcement to help users
retain information, provide clear and concise instructions for the user, use visual aids
such as images and diagrams to support your prompts, provide additional resources
for users who want to learn more, use a conversational tone to make your prompts
more approachable, encourage user interaction and engagement with your prompts,
use personalization to tailor your prompts to individual users, continually update
and refresh your prompts to keep them relevant, and most importantly, monitor the
performance of your prompts and make adjustments as needed.
It is important to note that optimizing prompts for language models like ChatGPT
is an ongoing process. As language models evolve and improve, and as user needs and
expectations change, it is important to continually monitor and update your prompts
to ensure that they remain effective and relevant most especially using data-driven
techniques. This can involve gathering data on how users interact with your prompts,
analyzing the responses generated by the language model, and using this information
to make data-driven decisions about how to improve your prompts. By continually
394 G. Marvin et al.

testing and refining your prompts, you can optimize their performance and achieve
better results.

3.2 Advanced Techniques for Prompt Engineering

Prompt engineering can be used to create effective prompts for a wide range of natural
language processing tasks, including Text Summarization, Question Answering,
Text Classification, Role Playing, Code Generation, Reasoning, Text Generation,
Text Translation, Sentiment Analysis, Named Entity Recognition, Text Comple-
tion, Dialog Generation, Paraphrasing, Text Simplification, Text-to-Speech, Speech-
to-Text, Image Captioning, Text-based Game Playing, Poetry Generation, Lyric
Generation, Story Generation, Joke Generation, Recipe Generation, Email Gener-
ation, Resume Generation, News Article Generation, Text-based Adventure Game
Playing, Text-based Puzzle Solving, Text-based Strategy Game Playing, Text-based
Simulation Game Playing, Text-based Interactive Fiction, Text-based Virtual Assis-
tant, Text-based Customer Service, Text-based Personal Shopping Assistant, Text-
based Personal Finance Assistant, and Text-based Personal Health Assistant among
others [1–30]. Accomplishing such tasks often requires basic and advanced prompt
engineering techniques. These kinds of techniques are explained in Table 2.
The prompt engineering techniques can be used to improve the performance of
language models like ChatGPT, allowing them to generate more coherent, relevant,
and sophisticated responses to user inputs. Besides those, there are some much more
advanced prompts engineering techniques with include;
Automatic Instruction Generation: It refers to the use of technology to auto-
matically generate instructions or prompts. This can be done in various contexts,
such as creating how-to guides or generating prompts for machine learning models.
One example of an advanced technique for automatic instruction generation tool
is the Automatic Prompt Engineer (APE), which was proposed for generating and
selecting instructions for Large Language Models. APE considers the instruction as
a “program” and optimizes it by searching through a pool of instruction candidates
suggested by a Large Language Model to maximize a chosen score function [37]. The
effectiveness of the chosen instruction is then assessed by measuring the zero-shot
performance of another Large Language Model that follows the selected instruc-
tion. Experiments on 24 NLP tasks showed that instructions automatically generated
using APE outperformed previous benchmarks and achieved better or similar perfor-
mance to instructions created by human annotators on 19 out of 24 tasks [38]. This
demonstrates the potential of automatic instruction generation as an advanced prompt
engineering technique.
Program Synthesis for Prompt Engineering: Classical program synthesis and
human approach prompt engineering is what inspires this technique. It entirely
relies on automatically generating instructions or prompts by using techniques from
program synthesis [38]. APE is still a good example for selected LLMs. Here, it
searches and optimized instruction as a program within a pool of LLMs while
Prompt Engineering in Large Language Models 395

Table 2 Prompt engineering techniques

Technique Description Most recommended large
language models (LLMs)
Few-shot prompts [19] It involves using a small number GPT-3, GPT-4
of carefully crafted prompts to
enable a language model to
perform a new task or adapt to a
new domain [19]
Chain-of-thought (CoT) Using a sequence of GPT-3, GPT-4
prompting [20] interconnected prompts to guide
the language model’s responses
in a coherent and logical manner
[20]
Self-consistency [21] Using prompts to encourage the GPT-3, GPT-4
language model to generate
responses that are consistent with
its previous responses [21]
Knowledge generation Using prompts to encourage the GPT-3, GPT-4
prompting [22] language model to generate new
knowledge or insights based on
its existing knowledge and
understanding of the world [23]
Reasoning and acting Using prompts to encourage the GPT-3, GPT-4
(ReAc) [24] language model to reason about a
given situation and generate
appropriate actions or responses
[24]
Contextual prompting [25] Providing additional context to Transformer-based models
the language model to help it such as GPT-3, GPT-4, and
generate more coherent and BERT
relevant responses [25]
Dynamic prompting [26] Dynamically adjusting the Transformer-based models
prompt based on the language such as GPT-3, GPT-4, and
model’s previous responses to BERT
improve its performance over
time [27]
Multi-modal prompting Using multiple modalities, such Multi-modal models such as
[28] as text and images, to provide CLIP and DALL-E
richer and more detailed prompts
to the language model [29]
Adversarial prompting [30] Using adversarial examples to Transformer-based models
test and improve the robustness such as GPT-3, GPT-4, and
of the language model’s BERT
responses to prompts [30]
Transfer learning Using transfer learning to adapt a Transformer-based models
prompting [31] pre-trained language model to such as GPT-3, GPT-4, and
new tasks or domains using BERT
carefully crafted prompts [31]
(continued)
396 G. Marvin et al.

Table 2 (continued)
Technique Description Most recommended large
language models (LLMs)
Meta-learning prompting Using meta-learning to train a Meta-learning models such as
[32] language model to quickly adapt MAML and reptile
to new tasks or domains using a
small number of carefully crafted
prompts [32]
Zero-shot prompting [33] Using carefully crafted prompts Transformer-based models
to enable a language model to such as GPT-3, GPT-4, and
perform tasks that it has not been BERT
explicitly trained on [33]
Active learning prompting Using active learning to Transformer-based models
[34] iteratively improve the such as GPT-3, GPT-4, and
performance of a language model BERT
by selecting the most informative
prompts for training [34]
Curriculum learning Using curriculum learning to Transformer-based models
prompting [35] gradually increase the difficulty such as GPT-3, GPT-4, and
of the prompts used to train a BERT
language model, allowing it to
learn more effectively [35]
Reinforcement learning Using reinforcement learning to Reinforcement learning
prompting [36] train a language model to models such as PPO and A2C
generate responses that
maximize a reward signal based
on the quality of its responses to
prompts [36]

continuously maximizing its score functions [38]. Still, zero-shot performance is

used measure the effectiveness of the prompt for overall evaluation within other
LLMs.
Program synthesis for prompt engineering and automatic instruction generation
is related but distinct concepts. Automatic instruction generation refers to the use of
technology to automatically generate instructions or prompts. This can be done in
various contexts, such as creating how-to guides or generating prompts for machine
learning models. Program synthesis for prompt engineering, on the other hand, refers
to the use of techniques from program synthesis to automatically generate instruc-
tions or prompts [39]. Program synthesis is fundamentally a field of research that
focuses on automatically generating programs from high-level specifications, such
as examples or natural language descriptions. In the context of prompt engineering,
program synthesis techniques are used to generate instructions or prompts that meet
certain criteria or specifications. The primary difference between the two concepts
is that automatic instruction generation is a broader term that encompasses various
techniques for generating instructions automatically, while program synthesis for
Prompt Engineering in Large Language Models 397

prompt engineering refers specifically to the use of program synthesis techniques for
this purpose.

3.3 Demonstration Tasks for Prompt Engineering

Few-shot learning: This technique enables in-context learning where demonstrations

are provided in the prompt to steer the model to better performance [19]. The demon-
strations serve as conditioning for subsequent examples where the model is expected
to generate a response.
Exercise 1: Experiment with few-shot learning by providing a series of messages
between the user and the assistant in the prompt as few-shot examples. Understand
how this technique improves accuracy and response grounding in LLMs.
Exercise 2: Execute few-shot learning for NLP tasks like text classification,
sentiment analysis, and language translation using LLMs.
Non-chat scenarios: This provides a better understanding of application of prompt
engineering can for specific tasks, e.g., answering questions on specific topics [40].
Exercise 1: Use ChatGPT or GPT-4 for a non-chat scenario task like generation
text that satisfies specific questions, e.g., “What is the capital city of Uganda?”
Exercise 2: Explore how prompt engineering can be applied to non-chat scenarios,
such as generating text for a specific purpose or answering questions on a specific
topic with Large Language Models.
Start with clear instructions: Experiment with providing clear instructions to the
model in the prompt to see how it affects the generated responses.
Exercise 1: Experiment with providing clear instructions to the model in the
prompt to see how it affects the generated responses. Fine tune the prompt to under-
stand how this technique can be used with alternative LLMs to improve accuracy
and grounding of given responses.
Exercise 2: Explore different techniques for providing clear instructions to the
model in the prompt when working with Large Language Models.
Explore different APIs: For Azure OpenAI GPT models, there are two separate
APIs where prompt engineering is important: the Chat Completion API and the
Completion API. Each API requires the input data to be structured in a specific way,
which affects the overall design of the prompt.
Exercise 1: Experiment with using different APIs for interacting with Azure
OpenAI GPT models. Explore how these APIs can be used with Large Language
Models to improve their accuracy and grounding of responses.
Exercise 2: Explore how different APIs require input data to be formatted differ-
ently and how this impacts overall prompt design when working with Large Language
Models.
Validate responses: Even when using prompt engineering effectively, it is crucial
to verify the responses generated by the models.
398 G. Marvin et al.

Exercise 1: Experiment with different techniques for validating responses gener-

ated by LLMs. Fine tune the prompts to comprehend the limitations of LLMs based
on the known strengths and weakness.
Exercise 2: Explore different techniques for validating responses generated by
LLMs when working with Large Language Models.

4 Applications, Tools, and Trends of Prompt Engineering

4.1 Applications and Tools

Since prompt engineering is a technique used to improve the performance of natural

language processing (NLP) models through providing better and more focused input,
there are several tools and applications available to facilitate this process for example;
LangChain, a library aimed at assisting in the development of applications that
combine large language models with other sources of computation or knowledge
[41]. [Link] which is a platform that helps build large language model applica-
tions as a series of prompted calls to external models [42]. OpenPrompt which is
a PyTorch-based library that offers a standard, adaptable, and expandable structure
for implementing the prompt-learning process. [43]. BetterPrompt, another test suite
for large language model prompts before pushing them to production [44]. Prompt
Engine which is an NPM utility library for creating and maintaining prompts for
large language models [45]. Promptify, a library aimed at assisting in developing a
pipeline for using large language model APIs in production [46]. TextBox 2.0 which
is a modern text generation library that uses Python and PyTorch to create a consistent
and standardized process for using pre-trained language models for text generation
[47]. ThoughtSource is a central, open resource, and community centered on data
and tools for chain-of-thought reasoning in large language models [48]. There is
also GPT Index. It is a very useful project constituting data structures utilized for
easing the usage of large external knowledge bases with LLMs [49, 50]. Moreover,
a community-driven initiative was also established to gather numerous prompts for
different cases. The community can be accessed via [Link]
fka/awesome-chatgpt-prompts.

4.2 Current Research and Future Trends of Prompt

Engineering

We have to acknowledge the fact that the rapid development of conversational AI has
underlined lots of emerging Natural Language Processing (NLP) innovations. One of
which are Large Language Models (LLMs). Such innovations have attracted attention
which has particularly facilitated the growth in size, capabilities, and transformative
Prompt Engineering in Large Language Models 399

potential of NLP driven technologies. We observe how such technologies have been
scaled to various domains like education and health care. It is important to note
that LLM operations involve creating and optimizing methods for input (prompts)
in order to obtain appropriate good quality output (responses). This interdependency
and need for good quality outputs across various domains of application is what
has inspired research initiatives and defined for us various research trend in prompt
engineering.
Among the inspired directions of prompt engineering research is the development
of better and more sophisticated domains specific prompt engineering tools and
techniques for interfacing with LLMs. The other trend is utilization of advanced
techniques in machine learning to automate the synthesis and optimization of prompts
for specific tasks. Opportunities are also opening up to integrate prompt engineering
in other NLP research areas like transfer learning and domain adoption. This focuses
on developing methods to adapt existing prompts to new domains or tasks. It also fits
into transfer of knowledge learned from one task to another prompt driven task. As
the world is working toward Society 5.0, a vision where every technology responsibly
connects and interacts with other technologies, it opens up research opportunities on
responsible utilization of prompt engineering for improved reliability and safety of
LLMs. Such research initiatives usually involve developing methods for detection
and mitigation of potential biases and errors in LLM inputs and outputs. It also
covers aspects for ensuring that LLMs behave in predictable, traceable, trustworthy,
and explainable means when applied in sensitive domains like health care.
Since prompt engineering is becoming a programming paradigm in conversational
AI, it opens up opportunities for utilizing prompts to improve explainability, inter-
pretability, and reliability of conversational AI applications most especially LLMs.
This can involve developing of methods for generating prompts that enable users
to understand how LLMs arrive to their outputs. This requires providing LLM
explainers for LLM output decisions. Building an equitable and well-connected
world drives attention on utilization of prompt engineering for efficiently and scal-
able deployment of LLMs. This opens up research opportunities on methods for
reducing computational costs of LLMs. Some these include how humans can enable
usage of large-scale datasets and distributed computing environments for LLMs.

4.3 Current Research and Future Trends of Large Language

Models (LLMs)

We are certain of increasing applications of LLMs across domains. This increase

comes with attention to the processes of developing larger and more powerful conver-
sational AI tools that accurately perform a variety of tasks that require human intel-
ligence. This will likely manifest as Conversational Artificial General Intelligence
(CAGI).
400 G. Marvin et al.

This opens up research trends building methods for developing, deployment,

and monitoring LLMs. It opens up research opportunities on equitable training and
utilization of LLMs among various social groups. Methods of transfer learning,
domain adaptation, and training LLMs on small amounts of data also open up. There
are inspired research trends to investigate reasoning, planning, and decision making
within various architectures of LLMs. For matters of generalizability and scalability,
the future and trend of research on LLMs needs to focus on combination of tech-
nologies (Multi-modeling) for NLP, computer vision, robotics, and quantum AI.
This requires development of optimal methods for integrating those technologies
for effective operations. The other trend in LLMs is Responsible Conversational AI,
this involves interpretability, explainability, inclusivity, trustworthiness, and respon-
sibility in modeling, development, deployment, and monitoring LLMs as conversa-
tional AI systems. Finally, human–machine interaction is another interest developing
among multiple disciplines who want to understand domain-specific human language
and behavior for effective communication and control of LLMs across domains.
Whereas LLMs are useful tools, they are still limited by correctness, lack of enter-
prise context, stale training data, limited controllability, and private data risks. These
make them perpetuate stereotypes and harm disadvantaged groups, they also spread
misinformation, outright disinformation and are still constrained by computing
power.

5 Conclusion

The potential and significance of prompt engineering on Large Language Models

(LLMs) and conversational AI systems is undeniable. With the impressive perfor-
mance demonstrated by LLMs on a wide range of NLP tasks in complex appli-
cations and use cases, we declare urgent and extra attention on prompt engineering
processes and procedures. Since prompt engineering is now a programming paradigm
for conversational AI, information seekers, AI engineers, and researchers need to get
excited about the new opportunities and challenges that come with developing more
powerful and capable Conversational AI systems.
By leveraging the latest advances in these fields, engineers and researchers can
build systems that can perform a wide range of tasks more effectively and effi-
ciently. Part of such tasks is building domain specific prompt engineering platfoms
based on responisble data standards and AI practices. These will be very impo-
rant for prompt engineering research, digital equality, coversational AI research and
industrial practices.
For ordinary users and information seekers, LLMs and prompt engineering have
the potential to significantly improve their interactions with Conversational AI
systems. By enabling AI systems to better understand and respond to human language
and behavior, LLMs and prompt engineering will make it easier for users to commu-
nicate with and control these systems. Therefore, active participation and staying
up-to-date with the latest advances in these fields, AI engineers, researchers, and
Prompt Engineering in Large Language Models 401

ordinary information seekers can all benefit from the exciting new opportunities
offered by LLMs and prompt engineering.

References

1. Brants T, Popat AC, Xu P, Och FJ, Dean J (2023) Large language models in machine translation.
[Link]. Online. Available: [Link] Accessed 01
May 2023
2. Du Y et al (2023) Guiding pretraining in reinforcement learning with large language models.
arXiv [Link]
3. Wang Y et al (2022) AdaMix: mixture-of-adaptations for parameter-efficient model tuning.
[Link]
4. Wei J et al (2022) Emergent abilities of large language models. arXiv [Link]
5. Oppenlaender J (2022) A taxonomy of prompt modifiers for text-to-image generation. arXiv
[Link]
6. White J et al (2023) A prompt pattern catalog to enhance prompt engineering with ChatGPT.
arXiv [Link]
7. Lo LS (2023) The CLEAR path: a framework for enhancing information literacy through
prompt engineering. J Acad Libr 49(4):102720
8. Short CE, Short JC (2023) The artificially intelligent entrepreneur: ChatGPT, prompt
engineering, and entrepreneurial rhetoric creation. J Bus Ventur Insights 19(e00388):e00388
9. Strobelt H et al (2023) Interactive and visual prompt engineering for ad-hoc task adaptation
with large language models. IEEE Trans Vis Comput Graph 29(1):1146–1156
10. Abukhalaf S, Hamdaqa M, Khomh F (2023) On codex prompt engineering for OCL generation:
an empirical study. arXiv [Link]
11. Oppenlaender J, Linder R, Silvennoinen J (2023) Prompting AI art: an investigation into the
creative skill of prompt engineering. arXiv [Link]
12. Chalkidis I (2023) ChatGPT may pass the bar exam soon, but has a long way to go for the
LexGLUE benchmark. arXiv [Link]
13. Johnson C, Rodríguez-Fernández N, Rebelo SM (2023) Artificial intelligence in music,
sound, art and design. In: 12th international conference, EvoMUSART 2023, held as part of
EvoStar 2023, Brno, Czech Republic, Apr 12–14, 2023, proceedings. Springer Nature, Cham,
Switzerland
14. Shtedritski A, Rupprecht C, Vedaldi A (2023) What does CLIP know about a red circle? Visual
prompt engineering for VLMs. arXiv [Link]
15. Polak MP, Morgan D (2023) Extracting accurate materials data from research papers with
conversational language models and prompt engineering—example of ChatGPT. arXiv [Link]
16. Busch K, Rochlitzer A, Sola D, Leopold H (2023) Just tell me: prompt engineering in business
process management. arXiv [Link]
17. Kumar K (2023) Geotechnical parrot tales (GPT): harnessing large language models in
geotechnical engineering. arXiv [Link]
18. Trautmann D, Petrova A, Schilder F (2022) Legal prompt engineering for multilingual legal
judgement prediction. arXiv [Link]
19. Ahmed T, Pai KS, Devanbu P, Barr ET (2023) Improving few-shot prompts with relevant static
analysis products. arXiv [Link]
20. Diao S, Wang P, Lin Y, Zhang T (2023) Active prompting with chain-of-thought for large
language models. arXiv [Link]
21. Taveekitworachai P, Abdullah F, Dewantoro MF, Thawonmas R, Togelius J, Renz J (2023)
ChatGPT4PCG competition: character-like level generation for science birds. arXiv [Link]
22. Kather JN, Ghaffari Laleh N, Foersch S, Truhn D (2022) Medical domain knowledge in domain-
agnostic generative AI. NPJ Digit Med 5(1):90
402 G. Marvin et al.

23. van Dis EAM, Bollen J, Zuidema W, van Rooij R, Bockting CL (2023) ChatGPT: five priorities
for research. Nature 614(7947):224–226
24. Yang Z et al (2023) MM-REACT: prompting ChatGPT for multimodal reasoning and action.
arXiv [Link]
25. Khattak MU, Rasheed H, Maaz M, Khan S, Khan FS (2022) MaPLe: multi-modal prompt
learning. arXiv [Link]
26. Wang B, Deng X, Sun H (2022) Iteratively prompt pre-trained language models for chain of
thought. arXiv [Link], pp 2714–2730
27. Liu P, Yuan W, Fu J, Jiang Z, Hayashi H, Neubig G (2023) Pre-train, prompt, and predict: a
systematic survey of prompting methods in natural language processing. ACM Comput Surv
55(9):1–35
28. Yang Z, Li Z, Zheng F, Leonardis A, Song J (2022) Prompting for multi-modal tracking. In:
Proceedings of the 30th ACM international conference on multimedia
29. Zhu J, Lai S, Chen X, Wang D, Lu H (2023) Visual prompt multi-modal tracking. arXiv [Link]
30. Maus N, Chao P, Wong E, Gardner J (2023) Adversarial prompting for black box foundation
models. arXiv [Link]
31. Wang Z, Panda R, Karlinsky L, Feris R, Sun H, Kim Y (2023) Multitask prompt tuning enables
parameter-efficient transfer learning. arXiv [Link]
32. Zhang H, Zhang X, Huang H, Yu L (2022) Prompt-based meta-learning for few-shot text
classification. In: Proceedings of the 2022 conference on empirical methods in natural language
processing, pp 1342–1357
33. Kojima T, Gu SS, Reid MM, Matsuo Y, Iwasawa Y (2022) Large language models are zero-shot
reasoners. arXiv [Link]
34. Köksal A, Schick T, Schütze H (2022) MEAL: stable and active learning for few-shot
prompting. arXiv [Link]
35. Lin J, Chen Q, Zhou J, Jin J, He L (2022) CUP: curriculum learning based prompt tuning for
implicit event argument extraction. arXiv [Link]
36. Zhang T, Wang X, Zhou D, Schuurmans D, Gonzalez JE (2022) TEMPERA: test-time
prompting via Reinforcement learning. arXiv [Link]
37. Zhou Y et al (2022) Steering large language models using APE
38. Zhou Y et al (2022) Large language models are human-level prompt engineers. arXiv [Link]
39. Austin J et al (2021) Program synthesis with large language models. arXiv [Link]
40. Sun K et al (2020) Adding chit-chat to enhance task-oriented dialogues. arXiv [Link]
41. Chase H (2023) Welcome to langchain—langchain 0.0.154. [Link]. Online. Available:
[Link] Accessed 01 May 2023
42. Dust—design and deploy large language models apps. [Link]. Online. Available: [Link]
[Link]/. Accessed 01 May 2023
43. “OpenPrompt,” [Link]. Online. Available: [Link] Accessed 01 May
2023
44. “The art & science of AI prompts,” The Art & Science of AI Prompts. Online. Available:
[Link] Accessed 01 May 2023
45. “[Link],” [Link]. Online. Available: [Link]
[Link]?traffic_id=GoDaddy_DLS&traffic_type=TDFS&utm_campaign=TDFS_G
oDaddy_DLS&utm_medium=sn_affiliate_click&utm_source=TDFS. Accessed 01 May 2023
46. “[Link],” [Link]. Online. Available: [Link] Accessed 01 May
2023
47. TextBox: TextBox 2.0 is a text generation library with pre-trained language models
48. ThoughtSource: A central, open resource for data and tools related to chain-of-thought
reasoning in large language models. Developed @ Samwald research group: [Link]
info/
49. G.-3 Demo, “GPT index,” [Link]. Online. Available: [Link]
index. Accessed 01 May 2023
50. “llamaindex (LlamaIndex),” [Link]. Online. Available: [Link]
ndex. Accessed 01 May 2023