Bridging AutoML and LLMs: Towards a Framework for Accessible and
Adaptive Machine Learning
Rafael Duque
1 a
, Cristina T
ˆ
ırn
˘
auc
˘
a
1 b
, Camilo Palazuelos
1 c
, Abraham Casas
2 d
,
Alejandro L
´
opez
2 e
and Alejandro P
´
erez
2 f
1
Department of Mathematics Statistics and Computer Science, University of Cantabria,
Avenida de los Castros S/N, Santander, Spain
2
Centro Tecnol
´
ogico CTC, Parque Cient
´
ıfico y Tecnol
´
ogico de Cantabria, Santander, Spain
Keywords:
Automated Machine Learning, Large Language Models.
Abstract:
This paper introduces a framework architecture that integrates Automated Machine Learning with Large Lan-
guage Models to facilitate machine learning tasks for non-experts. The system leverages natural language
processing to help users describe datasets, define problems, select models, refine results through iterative
feedback, and manage the deployment and ongoing maintenance of models in production environments. By
simplifying complex machine learning processes and ensuring the continued performance and usability of de-
ployed models, this approach empowers users to effectively apply machine learning solutions without deep
technical knowledge.
1 INTRODUCTION
Automated Machine Learning (AutoML) has
emerged as a transformative approach to democratize
access to machine learning (ML). By automating
the selection of models, hyperparameter tuning,
and feature engineering, AutoML systems allow
users with minimal expertise to leverage powerful
ML techniques (Karmaker et al., 2021). AutoML
solutions streamline the ML pipeline, reducing the
time and effort required to develop robust models
while maintaining or even improving their predictive
performance.
Large Language Models (LLMs) represent a sig-
nificant advancement in natural language processing
(NLP). Trained on massive datasets, these models
can generate human-like text, perform complex rea-
soning tasks, and adapt to a wide range of applica-
tions, from translation to summarization (Fan et al.,
2024). The flexibility of LLMs lies in their ability
to learn and generalize across diverse tasks with min-
a
https://orcid.org/0000-0001-8636-3213
b
https://orcid.org/0000-0002-7129-2237
c
https://orcid.org/0000-0003-4132-9550
d
https://orcid.org/0000-0002-7060-9298
e
https://orcid.org/0000-0002-3262-0605
f
https://orcid.org/0000-0002-4262-9275
imal fine-tuning, making them an invaluable tool for
both research and practical applications. Their perfor-
mance on benchmark tasks has pushed the boundaries
of what was previously thought achievable in NLP
(Chang et al., 2024).
The integration of LLMs into AutoML systems
has the potential to empower non-expert users, en-
abling individuals without a data science background
to effectively utilize AutoML tools. LLMs can en-
hance AutoML by interpreting natural language prob-
lem descriptions, suggesting suitable algorithms, and
explaining model outputs to users in an intuitive
manner (Tornede et al., 2023). Furthermore, LLMs
can assist in automating the creation of custom data
pipelines and model configurations based on specific
user requirements. This synergy between LLMs and
AutoML holds the potential to make ML even more
accessible and efficient, further closing the gap be-
tween technical complexity and practical usability.
Designing framework architectures that integrate
Large Language Models (LLMs) with AutoML to
meet the needs of non-expert users is a challenging
task. This involves ensuring that LLMs can under-
stand user instructions, suggest appropriate solutions,
and guide them through the process. At the same
time, the framework needs to deliver high-quality re-
sults while keeping the experience easy and acces-
sible for users from different fields. To address this
Duque, R., Tîrn
ˇ
auc
ˇ
a, C., Palazuelos, C., Casas, A., López, A. and Pérez, A.
Bridging AutoML and LLMs: Towards a Framework for Accessible and Adaptive Machine Learning.
DOI: 10.5220/0013448500003929
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 27th International Conference on Enterprise Information Systems (ICEIS 2025) - Volume 1, pages 959-964
ISBN: 978-989-758-749-8; ISSN: 2184-4992
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.
959

challenge, this article presents a proposal that aims to
consolidate the advancements made by the scientific
community in this area, offering a comprehensive ap-
proach to bridge existing gaps.
The structure of this article is as follows. Sec-
tion 2 reviews related work in the field of AutoML
and LLM integration, highlighting existing frame-
works and their limitations. In Section 3, we present
the architecture of our proposed framework, detail-
ing its design and key components. Section 4 pro-
vides a discussion of the framework’s strengths, in-
cluding its adaptability and usability for non-expert
users, as well as its potential impact compared to ex-
isting approaches. Finally, Section 5 concludes the
paper, summarizing the main contributions and out-
lining future directions for research and development.
2 RELATED WORK
Recent advancements in AutoML frameworks have
focused on enhancing usability and performance
through the integration of LLMs. Thus, AutoM3L
(Luo et al., 2024) stands out as a comprehensive sys-
tem designed for multimodal ML tasks, addressing
the limitations of earlier frameworks like AutoGluon.
It leverages LLMs to automate main steps such as
feature engineering, model selection, and pipeline as-
sembly, thus reducing the steep learning curve for
non-expert users. The framework achieves compet-
itive performance on various multimodal datasets,
showcasing the potential of LLMs to streamline com-
plex ML processes.
Aliro (Choi et al., 2023) presents a user-friendly
AutoML tool specifically designed for the biomedical
and healthcare domains. It combines an intuitive web
interface with the power of LLMs to assist users in
automating data analysis and ML tasks. By provid-
ing a conversational interface, Aliro allows users to
interact dynamically with their data, offering code ex-
ecution and model recommendations. This approach
significantly lowers the barriers for researchers with-
out programming expertise, enhancing accessibility
to advanced ML tools.
GizaML (Sayed et al., 2024) introduces a meta-
learning framework tailored for automated time-
series forecasting. It employs an LLM-based meta-
model to recommend optimal ML configurations
based on dataset characteristics. The framework’s
ability to handle complex tasks such as feature ex-
traction and hyperparameter optimization highlights
its effectiveness in real-time applications. GizaML’s
focus on time-series data addresses a critical need in
domains like energy and financial forecasting, where
traditional AutoML tools often fall short.
The integration of LLMs into AutoML frame-
works is further exemplified by systems like
SmartML(Maher and Sakr, 2019), TPOT (Le et al.,
2020), and JarviX (Liu et al., 2023). SmartML and
TPOT use meta-learning and genetic programming to
optimize pipelines, but they lack advanced user inter-
action features. JarviX addresses this gap by utiliz-
ing LLMs to guide users through a rule-based system
for data visualization and statistical analysis. JarviX
combines a vectorized domain knowledge repository
and fine-tuned models like Vicu
˜
na and GPT-4 to gen-
erate exploratory insights via text or voice input, cul-
minating in comprehensive reports. It further inte-
grates H2O-AutoML pipelines for customized model
training, enhancing its analytical depth and flexibility.
These advancements illustrate the significant
progress in making AutoML frameworks more acces-
sible to non-expert users through the integration of
LLMs. However, existing frameworks like AutoM3L
(Luo et al., 2024) focus primarily on automating core
steps such as feature engineering and model selec-
tion, while systems like GizaML (Sayed et al., 2024)
and JarviX (Liu et al., 2023) are domain-specific or
dependent on particular technologies. Our proposed
framework aims to address these gaps by offering a
more flexible, user-friendly approach that enables it-
erative model refinement and provides explainability,
all while being independent of specific technologies.
This flexibility and emphasis on explainability could
make this framework distinct from the existing solu-
tions. The following sections will present this frame-
work in detail, outlining its design and capabilities.
3 STRUCTURE OF THE
FRAMEWORK
This section presents the architecture of a user-
friendly framework that combines AutoML and
LLMs (see Figure 1), designed to empower users
without deep expertise in ML to leverage advanced
ML techniques for data analysis, problem defini-
tion, model training, and optimization. The frame-
work uses natural language processing (NLP) through
LLMs to guide the user through each step of the pro-
cess, simplifying the configuration and refinement of
ML workflows.
The framework begins with an Input Layer (see
Figure 1) where the user interacts with the system
through a natural language interface. The LLM acts
as a virtual assistant, helping the user describe the
dataset and define the problem to be solved. The user
uploads their dataset, which could be in various for-
ICEIS 2025 - 27th International Conference on Enterprise Information Systems
960

Figure 1: Architecture of the framework.
mats, such as CSV files, text documents, or databases.
The LLM processes the dataset and generates a sum-
mary that highlights the features, the types of data
(e.g., numerical, categorical, or text), and any poten-
tial relationships between variables. The LLM can
automatically detect missing values, identify outliers,
and suggest necessary preprocessing steps.
Using natural language, the user defines the prob-
lem they want to solve. For instance, the user might
describe a goal such as “predicting future sales” or
“classifying customer reviews.” The LLM translates
this high-level description into a specific ML task,
such as regression or classification, and prepares the
system to set up the AutoML process accordingly.
Once the problem is defined, the framework
moves into the Preprocessing Layer (see Figure 1).
At this stage, the system ensures the dataset is appro-
priately prepared for modeling. The LLM offers guid-
ance on the preprocessing steps that should be applied
to the data. It recommends handling missing data,
normalizing numerical features, encoding categorical
variables, or tokenizing text for NLP tasks. The LLM
provides explanations for each preprocessing tech-
nique, ensuring that users understand the purpose of
these steps. The system can automatically apply these
transformations based on the dataset’s characteristics,
allowing for a streamlined preparation process.
The next stage of the framework involves config-
uring the AutoML process, where the LLM plays a
main role in helping the user select appropriate set-
tings. For example, if the user is unsure whether to
select a decision tree or a neural network, the LLM
Bridging AutoML and LLMs: Towards a Framework for Accessible and Adaptive Machine Learning
961

can explain the advantages of each and recommend
which model might be most suitable for the task at
hand.
The AutoML system autonomously selects and
evaluates a range of ML models based on the config-
uration. The LLM helps users understand the selec-
tion process, explaining how AutoML identifies the
best models for the problem and adjusts hyperparame-
ters to maximize performance. Users are not required
to understand the underlying technicalities; the LLM
simplifies these concepts into user-friendly language.
The AutoML Module (see Figure 1) is the core
component of the framework, responsible for au-
tomating the ML workflow. This layer handles the
bulk of the computational processes, allowing users
to focus on higher-level decisions. Based on the con-
figuration, the AutoML module trains multiple ML
models using the dataset. This could include mod-
els like Random Forests, Support Vector Machines, or
neural networks. The system automatically splits the
data into training and validation sets, ensuring proper
model evaluation.
The AutoML system performs hyperparameter
optimization (see Figure 1) to fine-tune model pa-
rameters and improve the accuracy of the trained
models. Methods like grid search, random search,
or Bayesian optimization are applied to find the best
possible configuration for each model. After training,
the models are evaluated using appropriate metrics.
For classification tasks, common evaluation metrics
include accuracy, precision, recall, and F1-score. For
regression, metrics such as mean squared error (MSE)
or R-squared are used. The system generates an eval-
uation report that allows the user to assess model per-
formance.
Following the initial model evaluation, the frame-
work enters a feedback loop that integrates the LLM
to refine and optimize the process. The LLM reviews
the evaluation results and provides the user with an
easy-to-understand analysis of model performance.
For example, it can explain why a model performed
well or poorly based on certain features or hyperpa-
rameter settings. The LLM also offers visualizations
to help the user understand the model’s strengths and
weaknesses.
Based on the evaluation metrics, the LLM sug-
gests adjustments to improve model performance. For
instance, if the model performs poorly on certain sub-
sets of data, the LLM might recommend using addi-
tional features, applying different preprocessing tech-
niques, or altering the hyperparameter settings. The
user can ask the LLM specific questions like, “What
can we do to improve the model?” or “Why did the
model underperform?” The LLM interprets the user’s
queries and helps refine the AutoML pipeline accord-
ingly, guiding the user through the process of improv-
ing the model.
Once the LLM provides feedback and sugges-
tions, the framework enters an iterative process of
model reconfiguration. The AutoML system automat-
ically applies any suggested changes, such as retrain-
ing the model with different hyperparameters or addi-
tional data features. The LLM ensures that the user is
aware of all adjustments and their implications. The
training, evaluation, and refinement cycle continues
until the best model is identified, ensuring that the
user receives a model with optimal performance tai-
lored to their problem.
After the model has been trained and optimized,
the framework moves to the Deployment Layer (see
Figure 1), where the model is put into production. The
selected model is deployed into a production environ-
ment. The deployment process is automated, ensuring
that the user does not need to manage the intricacies
of model deployment. The LLM can help explain the
deployment process and provide guidance on how to
integrate the model into existing systems.
The system continuously monitors the model’s
performance in the real world. The LLM alerts the
user to potential issues, such as data drift or declining
accuracy, and suggests corrective actions. The LLM
can also assist in retraining the model if performance
degradation occurs over time.
Interpretability and explainability are critical for
understanding how models make decisions, especially
for non-experts. Thus, the Interpretability Layer
(see Figure 1) ensures that users can trust and un-
derstand their models. The LLM explains the rea-
soning behind the model’s predictions in a way that
is accessible to the user. For instance, it might say,
“The model predicted this outcome because there was
a high correlation between the values of feature X and
thos of the target variable”. This enhances the user’s
confidence in the model’s decisions. The LLM also
provides insights into the dataset and model predic-
tions, highlighting patterns in the data that the model
has identified. This helps the user gain a deeper un-
derstanding of the problem domain and the model’s
behavior.
The Resource Optimization Layer (see Figure
1) focuses on ensuring the framework operates effi-
ciently, even with large datasets or complex models.
The system manages computational resources, such
as GPUs and CPUs, to ensure cost-effective training
and inference. The LLM can suggest resource opti-
mization strategies to the user, helping them balance
performance with computational cost. The system is
designed to scale with the complexity of the task, en-
ICEIS 2025 - 27th International Conference on Enterprise Information Systems
962

suring that even large datasets can be processed effi-
ciently without compromising performance.
4 DISCUSSION
The framework offers a solution for empowering non-
expert users to leverage ML techniques. This section
discusses the key aspects of the framework architec-
ture, such as its user control, modularity, adaptability,
and its ability to offer explainability. Additionally, we
explore how the framework supports the iterative pro-
cess of ML, allowing users to refine their models and
improve performance continuously:
• User Control and Interaction. the framework
provides the user with clear steps, ensuring that
even those without technical expertise can engage
with ML processes effectively. The process be-
gins with the user describing the dataset and defin-
ing the problem. The LLM then assists in un-
derstanding the dataset and clarifying the prob-
lem, translating high-level descriptions into ML
tasks. This ensures the user retains control over
the workflow, while the system provides valuable
support in every step. For instance, once the user
defines the task, the LLM interprets it into a clear
ML objective, such as classification or regression.
This progression allows the user to guide the pro-
cess without needing to understand the intricacies
of the underlying algorithms.
• Modularity and Adaptability. the framework’s
various components, such as dataset understand-
ing, problem definition, preprocessing, and model
selection, work together in an adaptable system.
As the user proceeds through the workflow, the
system applies modular AutoML tools and adjusts
the configuration automatically based on user in-
puts. Each layer of the process, from data pre-
processing to model evaluation, is adaptable to
different types of datasets and problem domains.
Whether the user is dealing with time-series data,
images, or text, the framework adapts its methods
and tools to fit the task at hand, allowing users to
solve a broad range of problems without needing
to switch between different technologies or tools.
• Technology Independence. the framework’s de-
sign makes it independent of specific ML tech-
nologies or platforms. While the AutoML compo-
nent utilizes various ML models (decision trees,
neural networks, etc.), the user interacts with the
system through natural language and does not
need to worry about the technical details of the
underlying tools. The LLM provides an abstrac-
tion layer, allowing the user to remain agnostic
to the specifics of the AutoML tools or technolo-
gies. This independence from technology means
that the framework can easily incorporate new ad-
vancements in ML, ensuring the system remains
up-to-date without disrupting the user experience.
• Explainability and Interpretability. a major
challenge in ML is ensuring the results are under-
standable to non-experts. The proposal addresses
this challenge by providing explainability through
its LLM interface. The LLM explains model pre-
dictions and offers insights into how different fea-
tures influenced the outcome. This allows the user
to better understand the model’s behavior and gain
confidence in the results. For instance, when a
model makes a prediction, the LLM can explain
why certain features were important in driving the
decision. This transparent communication is cru-
cial for helping users make informed decisions
about further improving the model or selecting the
best one for deployment. In this way, the frame-
work enables the user to trust and interpret the re-
sults, which is especially important when the user
lacks the expertise to dive into the technical as-
pects of the model.
• Iterative Refinement and User Feedback. one
of the core features of the framework is the contin-
uous feedback loop that allows users to refine their
models iteratively. The LLM helps the user under-
stand why certain models performed better than
others and makes specific suggestions to improve
performance. The iterative cycle of training, eval-
uation, feedback, and reconfiguration ensures that
the model improves over time. The framework’s
design supports ongoing refinement based on user
interactions.
5 CONCLUSIONS
This work presents an architecture for a framework
that offers an accessible solution for users without
deep expertise in ML. By allowing users to engage
in tasks like dataset description, problem definition,
and model refinement through natural language, the
framework simplifies the process of ML. The integra-
tion of AutoML with LLMs makes it possible for non-
experts to navigate complex workflows with ease,
bridging the gap between advanced ML techniques
and user-friendly interfaces.
The framework also ensures iterative improve-
ment and clarity at each step, allowing users to re-
fine their models based on feedback from the system.
Bridging AutoML and LLMs: Towards a Framework for Accessible and Adaptive Machine Learning
963

By continuously guiding users through each stage of
the process, the system fosters a deeper understanding
of ML principles while enabling effective decision-
making without requiring advanced technical knowl-
edge.
Future work will focus on building ML models
and experimenting with the proposal to assess its ef-
fectiveness in real-world scenarios. This will include
testing the framework with diverse datasets and prob-
lem types to ensure its scalability and adaptability.
Moreover, LLMs may have limitations, such as high
computational costs, biases, and the potential for gen-
erating incorrect information. These issues were not
addressed in this paper but could be considered in fu-
ture work.
ACKNOWLEDGEMENTS
This work has been partially supported by grants
PID2022-139237NB-I00 and PID2023-146243OB-
I00 funded by MICIU/AEI/10.13039/501100011033
and by “ERDF/EU”. This research was also partially
developed in the project FUTCAN-2023/TCN/018
that was co-financed from the European Regional
Development Fund through the FEDER Operational
Program 2021-2027 of Cantabria through the line of
grants “Aid for research projects with high industrial
potential of excellent technological agents for indus-
trial competitiveness TCNIC”.
REFERENCES
Chang, Y., Wang, X., Wang, J., Wu, Y., Yang, L., Zhu, K.,
Chen, H., Yi, X., Wang, C., Wang, Y., et al. (2024). A
survey on evaluation of large language models. ACM
Transactions on Intelligent Systems and Technology,
15(3):1–45.
Choi, H., Moran, J., Matsumoto, N., Hernandez, M. E., and
Moore, J. H. (2023). Aliro: an automated machine
learning tool leveraging large language models. Bioin-
formatics, 39(10):btad606.
Fan, L., Li, L., Ma, Z., Lee, S., Yu, H., and Hemphill, L.
(2024). A bibliometric review of large language mod-
els research from 2017 to 2023. ACM Transactions on
Intelligent Systems and Technology, 15(5):1–25.
Karmaker, S. K., Hassan, M. M., Smith, M. J., Xu, L.,
Zhai, C., and Veeramachaneni, K. (2021). Automl to
date and beyond: Challenges and opportunities. ACM
Computing Surveys (CSUR), 54(8):1–36.
Le, T. T., Fu, W., and Moore, J. H. (2020). Scaling
tree-based automated machine learning to biomedical
big data with a feature set selector. Bioinformatics,
36(1):250–256.
Liu, S.-C., Wang, S., Lin, W., Hsiung, C.-W., Hsieh, Y.-
C., Cheng, Y.-P., Luo, S.-H., Chang, T., and Zhang,
J. (2023). Jarvix: A llm no code platform for tab-
ular data analysis and optimization. arXiv preprint
arXiv:2312.02213.
Luo, D., Feng, C., Nong, Y., and Shen, Y. (2024). Autom3l:
An automated multimodal machine learning frame-
work with large language models. In Proceedings of
the 32nd ACM International Conference on Multime-
dia, pages 8586–8594.
Maher, M. M. M. Z. A. and Sakr, S. (2019). Smartml: A
meta learning-based framework for automated selec-
tion and hyperparameter tuning for machine learning
algorithms. In EDBT: 22nd International conference
on extending database technology.
Sayed, E., Maher, M., Sedeek, O., Eldamaty, A., Kamel,
A., and El Shawi, R. (2024). Gizaml: A collabora-
tive meta-learning based framework using llm for au-
tomated time-series forecasting. In EDBT, pages 830–
833.
Tornede, A., Deng, D., Eimer, T., Giovanelli, J., Mohan,
A., Ruhkopf, T., Segel, S., Theodorakopoulos, D.,
Tornede, T., Wachsmuth, H., et al. (2023). Automl
in the age of large language models: Current chal-
lenges, future opportunities and risks. arXiv preprint
arXiv:2306.08107.
ICEIS 2025 - 27th International Conference on Enterprise Information Systems
964