A Collaborative Approach to Multimodal Machine Translation: VLM
and LLM
Amulya Ratna Dash and Yashvardhan Sharma
Birla Institute of Technology and Science, Pilani, Rajasthan, India
Keywords:
Natural Language Processing, Multimodal Machine Translation, Large Language Models, Image Captioning.
Abstract:
With advancements in Large Language Models (LLMs) and Vision Language Pretrained Models (VLMs),
there is a growing need to evaluate their capabilities and research methods to use them together for vision
language tasks. This study focuses on using VLM and LLM collaboratively for Multimodal Machine Transla-
tion (MMT). We finetune LLaMA-3 to use provided image captions from VLMs to disambiguate and generate
accurate translations for MMT tasks. We evaluate our novel approach using the German, French and Hindi
languages, and observe consistent translation quality improvements. The final model shows an improvement
of +3 BLEU score against the baseline and +4 BLEU score against the state-of-the-art model.
1 INTRODUCTION
Machine Translation (MT) is a Natural Language Pro-
cessing (NLP) task in which text from a source lan-
guage is translated to another target language while
preserving the semantics and required terminology.
Classical machine translation systems only perceive
textual information, ignoring useful information from
visual modalities such as images and videos. Mul-
timodal Machine Translation (MMT) is the process
where other modalities are used to improve the quality
of the language translation. The most popular modal-
ity in MMT is a visual clue or image. The visual
context helps disambiguate a few words by provid-
ing additional information along with source textual
context. MMT can be extensively adopted for di-
verse practical applications, including subtitle trans-
lations considering image/video along with original
source language movie script, and the translation of
product descriptions and reviews in e-commerce plat-
forms from original product listing language to pre-
ferred language of customer. Consequently, there is
an increasing focus on research in MMT.
In recent years, Large Language Models (LLMs)
have become increasingly widespread. Models like
GPTs (Brown et al., 2020), LLaMA (Touvron et al.,
2023), and others have shown remarkable perfor-
mance in various NLP tasks such as text generation,
question answering, and summarization. Vision Lan-
guage Models (VLMs) understand and generate lan-
guage in the context of visual information. These
models leverage large datasets of paired images and
text to jointly learn the visual and linguistic relation-
ship. VLMs have shown steady improvement in tasks
such as image captioning and visual question answer-
ing(VQA).
In this work, we propose a collaborative approach
to Multimodal Machine Translation, using Vision
Language Pretrained Models (VLMs) and Large Lan-
guage Models (LLMs). By integrating VLMs into our
methodology, we propose to utilize their ability to of-
fer enhanced contextual insights, thus boosting the ef-
fectiveness of Multimodal Machine Translation using
LLMs. Our findings show that fine-tuning LLMs with
extremely limited training data and computational re-
source can also help improve the translation accuracy
for both high resource and low resource languages.
The rest of the paper is organized as follows. Sec-
tion 2 presents the review of related works. The
methodology and dataset are briefly described in Sec-
tion 3. Sections 4 and 5 detail the experiments and
results, followed by the conclusion and future scope
in Section 6.
2 RELATED WORK
2.1 Multimodal Machine Translation
Most of the multimodal machine translation (MMT)
work are based on Multi30K (Elliott et al., 2016)
1412
Dash, A. R. and Sharma, Y.
A Collaborative Approach to Multimodal Machine Translation: VLM and LLM.
DOI: 10.5220/0013379900003890
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 17th International Conference on Agents and Artificial Intelligence (ICAART 2025) - Volume 3, pages 1412-1418
ISBN: 978-989-758-737-5; ISSN: 2184-433X
Proceedings Copyright © 2025 by SCITEPRESS – Science and Technology Publications, Lda.

dataset. Initial approaches focused on using visual
features in sequence-to-sequence models based on
Recurrent Neural Networks (RNNs) and Transform-
ers (Vaswani, 2017) architecture. (Calixto et al.,
2017) proposed a doubly-attentive model for both
source text and global image features. Gated Fusion
showed a mechanism to use image features in cross-
attention of decoder. (Libovick
`
y and Helcl, 2017)
used flat and hierarchical approach to combine the at-
tention mechanism for multimodal translation. Most
of the research was benchmarked on English to Ger-
man and French evaluation datasets.
Parida et al., 2019 proposed Hindi Visual Genome
dataset, a subset of the Visual Genome (Krishna et al.,
2017) dataset for multimodal translation between En-
glish and Hindi, allowing researchers to experiment
and benchmark using a new language. An adversarial
study of MMT showed replacing/removing the asso-
ciated image does not hamper the MMT output, which
suggests that source text is sufficient to generate tar-
get text semantically close to reference evaluation
text (Elliott, 2018). CoMMuTe (Futeral et al., 2022)
dataset helps to better evaluate MMT systems, as it
contains lexically ambiguous sentences and only the
associated image helps disambiguate. In our work,
we evaluate our MMT approach using CoMMuTe and
HVG datasets.
2.2 Pretrained Language Models and
MMT
mBART (Chipman et al., 2022), a denoising auto-
encoder pre-trained on monolingual corpora in mul-
tiple languages shows its effectiveness for machine
translation(MT). Zhu et al., 2020 used BERT to im-
prove neural machine translation, by fusing sentence
representations from BERT with each encoder and
decoder layer. CLIP (Bordes et al., 2024) model
learns to perform multiple computer vision tasks
like image captioning during pretraining, using con-
trastive language image pretraining approach. BLIP
(Li et al., 2022) model uses the vision language pre-
training approach which improves both vision lan-
guage understanding and generation tasks. BLIP 2
(Li et al., 2023) further improves the accuracy on
vision-language tasks by reducing the image and text
modality gap with a Querying Transformer. Florence-
2 (Yuan et al., 2021) is a vision-language model which
takes text instructions and generates textual results
based on the image. The performance of Florence-2-L
model(0.77B params) on captioning and visual ques-
tion answering tasks are comparable to other much
larger models like BLIP-2 and Flamingo (Alayrac
et al., 2022).
Futeral et al., 2022 propose a multimodal MT
model VGAMT based on mBART. VGAMT adds
bottleneck adapters (Houlsby et al., 2019), and lin-
ear visual projection layers to the finetuned mBART,
uses MDETR (Misra et al., 2021) and CLIP for lo-
cal and global image features, respectively. Our work
uses BLIP-2 and Florence to generate image captions,
followed by adapting a large language model to con-
dition upon the generated image caption while per-
forming the translation task.
2.3 Large Language Models and MT
The machine translation capability of LLMs has been
studied by multiple researchers. Xu et al. (2023)
proposed a paradigm shift in machine translation
which adapts LLaMA for translation using two steps,
continued pretraining on monolingual data and fine-
tuning using bilingual data. Tower LLM (Alves
et al., 2024) is a multilingual LLM for translation-
related tasks which also adapts LLaMA using two
steps, uses monolingual and parallel data in first step
and finetunes on instructions for multiple tasks like
translation, automatic post editing (APE), named en-
tity recognition (NER) and grammatical error correc-
tion(GEC) in the second step. In our work, we focus
on adapting LLaMA for multimodal machine transla-
tion via single-step fine-tuning using extremely lim-
ited amount of finetuning data tuples(image caption
and bilingual text data).
3 METHODOLOGY
3.1 Models
To investigate the collaborative approach to Multi-
modal Machine Translation, we require two cate-
gories of models: an Image-to-Text Model and a
Large Language Model. The Image-to-Text model
is used to generate the caption of the image associ-
ated with the source text. The Large Language Model
is used to translate the source language text into tar-
get language text, using the description of the image
generated by Image-to-Text model as context. BLIP-
2 and Florence-2 are used as Image-to-Text models
and LLaMa 3 as the Large Language Model. BLIP
2 offers multiple model variants based on underly-
ing backbone model and model size, for image en-
coder (ViT-L or ViT-g) and for text generation (OPT
2.7B, OPT 6.7B, FlanT5 XL, FlanT5 XXL). LLaMa
3 offers two model variants based on model size, a 2
billion and 8 billion parameters model. The small-
est model variant was selected so that with limited
A Collaborative Approach to Multimodal Machine Translation: VLM and LLM
1413

resource we can perform efficient finetuning and in-
ference.
We use BLIP-2 (VIT-L OPT-2.7B) model, which
uses OPT-2.7B model along with a CLIP-like image
encoder and a Querying Transformer to generate text
conditioned on image and optional text. LLaMa 3 8B
1
model was shortlisted as it focuses on being multilin-
gual with high quality non-English data in the train-
ing dataset, resulting in impressive zero-shot transla-
tion performance. LLaMA 3 8B model was finetuned
using the LoRA (Hu et al., 2021) parameter efficient
fine-tuning technique (PEFT).
3.2 Languages and Evaluation Data
To evaluate our approach on Multimodal machine
translation, we consider 3 languages: German, French
and Hindi. CoMMuTE, a Contrastive Multilingual
Multimodal Translation Evaluation dataset of lexi-
cally ambiguous sentences whose correct translation
requires the context from associated image. Mul-
timodal machine translation test sets like Test2016,
Test2017, MSCOCO, HVG are some of the other
available evaluation datasets. CoMMuTE (300 seg-
ments) was considered as the evaluation dataset for
German and French tasks, as other evaluation datasets
have relatively lower percentage of ambiguous exam-
ples, thus diluting the evaluation score. HVG Chal-
lenge set (1400 segments) was used as an evaluation
dataset for the Hindi task.
3.3 Evaluation Metrics
ChrF (Popovi
´
c, 2015), COMET-22 (Rei et al., 2022)
and BLEU (Papineni et al., 2002) score were used
as evaluation metrics. ChrF evaluates translations at
character level which make it more robust as com-
pared to token level metrics. Crosslingual Optimized
Metric for Evaluation of Translation 2022 (COMET
22) is an embedding-based evaluation metric, which
considers context and semantics of the text while as-
signing scores. BLEU score is calculated by compar-
ing the n-gram overlap between the system and the
reference texts.
1
https://huggingface.co/meta-llama/
Meta-Llama-3-8B-Instruct
4 EXPERIMENTS
4.1 Tasks
We consider three multimodal translation tasks: (i)
English → German (ii) English → French (iii) En-
glish → Hindi. First, we create a text only translation
baseline using LLaMa 3 model. Secondly, we per-
form multimodal machine translation using our cas-
caded approach, where we generate image descrip-
tion using BLIP-2 model and use it as context for
translation using LLaMa 3 model. Finally, we use
our finetuned LLaMA 3 model along with image cap-
tions generated from BLIP-2 and Florence-2 models
to perform multimodal machine translation.
4.2 Data
The finetuning dataset consists of German, French
and Hindi portions. We use the val split of Multi30K
dataset for German and French, and the dev split
of the HindiVisualGenome (HVG) dataset for Hindi.
The context (in English) part of our training dataset is
generated using the BLIP-2 model for the image asso-
ciated with each record. Then we process the context
phrase and bilingual text data into a chat instruction
template as required for finetuning LLaMa 3 model.
The template used for training data is available in Fig-
ure 1. The distribution of processed dataset is shown
in Table 1.
Table 1: Composition of dataset used for finetuning LLM.
Language Pair Train Val
English - German 900 100
English - French 900 100
English - Hindi 900 100
Total 2700 300
4.3 Training Setup
The model was finetuned for multilingual multimodal
machine translation using 2700 training records and
300 validation records. The target modules for LoRA
finetuning are Query, Key, Value, Output, Gate, Up
and Down projections. The hyperparameters are de-
tailed in Table 2.
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
1414

Figure 1: Template of training dataset.
Table 2: Training hyperparameters.
Parameters Values
Learning rate 0.0003
Train batch size 8
Eval batch size 4
Seed 3407
Gradient accumulation 8
Total train batch size 64
Optimizer Adam(Kingma,
2014)
Adam β
1
0.9
Adam β
2
0.999
Adam ε 1e-08
LR scheduler type cosine
LR warmup steps 0.01
Num of Epochs 2
Num of GPU 1 (A100 40GB)
LoRA Rank 16
LoRA α 16
LoRA Bias None
5 RESULTS & ANALYSIS
5.1 Results
We evaluated vanilla LLaMA 3 for Text only transla-
tion and used the scores as baseline result. The com-
parative scores from cascaded approach of VLM and
LLM (vanilla and finetuned) can be found in Table 3
for English → {German, French} and in Table 4 for
English → Hindi . We compared the state-of-the-art
BLEU scores reported by the authors of VGAMT on
CoMMuTE dataset with our best model as shown in
Table 5.
• CoMMuTe Eval set: Our model MMT-LLaMa-
3 when augmented with caption from Florence
model during inference received 5 best scores out
of 3 metrics each for English → German and En-
glish → French multimodal translations. COMET
score improved by +4 and +3.4 points, and ChrF
score boosted by +3.76 and +4.27 absolute points
for English → German and English → French re-
spectively.
MMT-LLaMa-3 + BLIP collaborative method
outperformed existing methods by a margin of
+4 for both English → {German, French} mul-
timodal translations, with only 2700 training data
and less than 15 minutes of GPU training time.
• HVG Challenge Eval set: Our model MMT-
LLaMa-3 + BLIP collaborative approach outper-
formed the MMT-LLaMa-3 + Florence-2 collab-
orative approach by +3.25 ChrF points. How-
ever, the clear winner on the multimodal English
→ Hindi translation was vanilla LLaMA 3 (text
translation without any hint from the associated
image) based on average scores.
5.2 Analysis
• German and French are from the Germanic and
Romance language family, and belong to Indo-
European languages. We observe almost similar
gains of +4 points for both German and French.
This can be attributed to the high resourceful-
ness of both languages and maybe to usage of an
equal number (around 900) of training data seg-
ments during the fine-tuning process. In addition,
the English source segments from the CoMMuTE
datasets overlap almost completely.
• The performance of the finetuned model on the
HVG Challenge set for English → Hindi multi-
modal translation was not aligned with our hy-
pothesis. The observation from analyzing few
individual translations are that majority of the
source segments in HVG Challenge set are of
shorter length relatively, which makes it easier to
A Collaborative Approach to Multimodal Machine Translation: VLM and LLM
1415

Table 3: MultiModal Translation performance for English → {German, French} on CoMMuTE dataset.
German French
Model BLEU COMET ChrF BLEU COMET ChrF
LLaMa 3 30.52 78.49 50.83 33.02 77.63 53.25
LLaMa 3 + BLIP 32.32 81.95 53.96
35.76 80.73 57.46
MMT-LLaMa-3 (Ours) + BLIP 33.3 82.36 54.31 36.28 80.89 56.82
MMT-LLaMa-3 (Ours) + Florence 32.79 82.47 54.59 36.69 81.04 57.52
Table 4: MultiModal Translation performance for English
→ Hindi on HVG Challenge dataset.
Model BLEU COMET ChrF
LLaMa 3 23.27 73.59 48.79
LLaMa 3 + BLIP 16.83 69.08 43.7
MMT-LLaMa-3
(Ours) + BLIP
11.5 72.5 43.31
MMT-LLaMa-3
(Ours) + Florence
9.31 71.14 40.06
Table 5: Comparison of BLEU score on English →
{German, French} on CoMMuTE dataset.
Model German French
VGAMT 29.3 32.2
MMT-LLaMa-3 +
BLIP (Ours)
33.3 36.28
translate directly. BLIP generated very short cap-
tions as compared to Florence, thus the shorter
context helped get better scores for translation as
the attention span reduced due to short and to-the-
point caption used as context. Few possible rea-
sons which we have not validated extensively may
be related to less image dependency and ambigu-
ity coverage in the evaluation dataset.
• Sample inference examples are shown in Table 6
and 7.
• In Table 6, the words “bow” are disambiguated
with the help of the image. The image shows a red
bow sitting on a table and a young boy holding a
bow and arrow. This context helps clarify that the
first “bow” refers to a decorative bow, while the
second “bow” refers to a weapon used for archery.
The image context facilitated the German transla-
tions, with “Schleier” for the decorative bow and
“Bogen” for the archery bow. In the first image, if
the VLM would have generated the caption with
word ‘red veil’ instead of ‘red bow’, the German
translation using our finetuned model would be
‘Schleife’.
• In Table 7, the words “chopper” are disam-
biguated with the help of the image. The image
shows a blue motorcycle parked in a parking lot
next to other motorcycles and a helicopter flying
in the sky over a house. This context helps clar-
ify that the first “chopper” refers to a motorcycle,
while the second “chopper” refers to a helicopter.
The French translation accurately reflects this dis-
tinction, with “chopper” for the motorcycle and
“h
´
elicopt
`
ere” for the helicopter.
6 CONCLUSION AND FUTURE
WORK
The multilingual ability of LLMs are increasing
rapidly with the release of new LLMs. These LLMs
being trained on enormous amounts of web scale data
and using extremely powerful GPU clusters; they
should be utilized for novel use cases. Most of the
times, in-context learning or adapter based finetuning
of LLMs helps to get acceptable accuracy for some
NLP problems like machine translation with limited
data and GPU power.
Our study concludes that using a collaborative ap-
proach of using a pre-trained small vision language
foundation models (VLM) and multilingual large lan-
guage models (LLM) jointly can provide a resource-
efficient solution to multimodal machine translation
of both low resource and high resource languages. An
improvement of +4 in the ChrF score and +3 points
for BLEU and COMET score was achieved compared
to the baseline score for the German and French trans-
lation tasks.
In the future, we would experiment the multi-
modal machine translation problem with additional
language pairs and perform a comparative evaluation
of multiple large language models on diverse evalua-
tion dataset to develop a generalized highly accurate
MMT system.
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
1416

Table 6: Sample multimodal translation of English → German, where Caption is used as context by the finetuned model as
visual hint.
Image
Caption: A red bow sitting on top of a table. A young boy holding a bow and arrow in
his hand.
Source Text: Hand me that bow. Hand me that bow.
Translated Text: Gib mir den Schleier. Gib mir jenen Bogen.
Reference
Translation:
Gib mir die Schleife. Gib mir den Bogen.
Table 7: Sample multimodal translation of English → French, where Caption is used as context by the finetuned model as
visual hint.
Image
Caption: A blue motorcycle parked in a parking lot
next to other motorcycles.
A helicopter flying in the sky over a house.
Source text: My husband bought a chopper recently. My husband bought a chopper recently.
Translated text: Mon mari a achet
´
e un chopper r
´
ecemment. Mon mari a achet
´
e un h
´
elicopt
`
ere
r
´
ecemment.
Reference
Translation:
Mon mari a achet
´
e une moto r
´
ecemment. Mon mari a achet
´
e un h
´
elicopt
`
ere
r
´
ecemment.
REFERENCES
Alayrac, J.-B., Donahue, J., Luc, P., Miech, A., Barr,
I., Hasson, Y., Lenc, K., Mensch, A., Millican, K.,
Reynolds, M., et al. (2022). Flamingo: a visual lan-
guage model for few-shot learning. Advances in neu-
ral information processing systems, 35:23716–23736.
Alves, D. M., Pombal, J., Guerreiro, N. M., Martins, P. H.,
Alves, J., Farajian, A., Peters, B., Rei, R., Fernandes,
P., Agrawal, S., et al. (2024). Tower: An open mul-
tilingual large language model for translation-related
tasks. arXiv preprint arXiv:2402.17733.
Bordes, F., Pang, R. Y., Ajay, A., Li, A. C., Bardes, A.,
Petryk, S., Ma
˜
nas, O., Lin, Z., Mahmoud, A., Jayara-
man, B., et al. (2024). An introduction to vision-
language modeling. arXiv preprint arXiv:2405.17247.
Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D.,
Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G.,
Askell, A., et al. (2020). Language models are few-
shot learners. Advances in neural information pro-
cessing systems, 33:1877–1901.
Calixto, I., Liu, Q., and Campbell, N. (2017). Doubly-
attentive decoder for multi-modal neural machine
translation. In Proceedings of the 55th Annual Meet-
ing of the Association for Computational Linguistics
(Volume 1: Long Papers), pages 1913–1924.
Chipman, H. A., George, E. I., McCulloch, R. E., and Shiv-
ely, T. S. (2022). mbart: multidimensional monotone
bart. Bayesian Analysis, 17(2):515–544.
Elliott, D. (2018). Adversarial evaluation of multimodal
machine translation. In Riloff, E., Chiang, D., Hock-
enmaier, J., and Tsujii, J., editors, Proceedings of
the 2018 Conference on Empirical Methods in Natu-
ral Language Processing, pages 2974–2978, Brussels,
A Collaborative Approach to Multimodal Machine Translation: VLM and LLM
1417

Belgium. Association for Computational Linguistics.
Elliott, D., Frank, S., Sima’an, K., and Specia, L. (2016).
Multi30k: Multilingual english-german image de-
scriptions. In Proceedings of the 5th Workshop on
Vision and Language, pages 70–74.
Futeral, M., Schmid, C., Laptev, I., Sagot, B., and Bawden,
R. (2022). Tackling ambiguity with images: Improved
multimodal machine translation and contrastive eval-
uation. arXiv preprint arXiv:2212.10140.
Houlsby, N., Giurgiu, A., Jastrzebski, S., Morrone, B.,
De Laroussilhe, Q., Gesmundo, A., Attariyan, M., and
Gelly, S. (2019). Parameter-efficient transfer learning
for nlp. In International conference on machine learn-
ing, pages 2790–2799. PMLR.
Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang,
S., Wang, L., and Chen, W. (2021). Lora: Low-rank
adaptation of large language models. arXiv preprint
arXiv:2106.09685.
Kingma, D. P. (2014). Adam: A method for stochastic op-
timization. arXiv preprint arXiv:1412.6980.
Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata,
K., Kravitz, J., Chen, S., Kalantidis, Y., Li, L.-
J., Shamma, D. A., et al. (2017). Visual genome:
Connecting language and vision using crowdsourced
dense image annotations. International journal of
computer vision, 123:32–73.
Li, J., Li, D., Savarese, S., and Hoi, S. (2023). Blip-
2: Bootstrapping language-image pre-training with
frozen image encoders and large language models. In
International conference on machine learning, pages
19730–19742. PMLR.
Li, J., Li, D., Xiong, C., and Hoi, S. (2022). Blip:
Bootstrapping language-image pre-training for unified
vision-language understanding and generation. In In-
ternational conference on machine learning, pages
12888–12900. PMLR.
Libovick
`
y, J. and Helcl, J. (2017). Attention strategies for
multi-source sequence-to-sequence learning. In Pro-
ceedings of the 55th Annual Meeting of the Associa-
tion for Computational Linguistics (Volume 2: Short
Papers), pages 196–202.
Misra, I., Girdhar, R., and Joulin, A. (2021). An end-to-end
transformer model for 3d object detection. In Pro-
ceedings of the IEEE/CVF international conference
on computer vision, pages 2906–2917.
Papineni, K., Roukos, S., Ward, T., and Zhu, W.-J. (2002).
Bleu: a method for automatic evaluation of machine
translation. In Isabelle, P., Charniak, E., and Lin, D.,
editors, Proceedings of the 40th Annual Meeting of
the Association for Computational Linguistics, pages
311–318, Philadelphia, Pennsylvania, USA. Associa-
tion for Computational Linguistics.
Parida, S., Bojar, O., and Dash, S. R. (2019). Hindi vi-
sual genome: A dataset for multi-modal english to
hindi machine translation. Computaci
´
on y Sistemas,
23(4):1499–1505.
Popovi
´
c, M. (2015). chrf: character n-gram f-score for auto-
matic mt evaluation. In Proceedings of the tenth work-
shop on statistical machine translation, pages 392–
395.
Rei, R., De Souza, J. G., Alves, D., Zerva, C., Farinha,
A. C., Glushkova, T., Lavie, A., Coheur, L., and Mar-
tins, A. F. (2022). Comet-22: Unbabel-ist 2022 sub-
mission for the metrics shared task. In Proceedings
of the Seventh Conference on Machine Translation
(WMT), pages 578–585.
Touvron, H., Martin, L., Stone, K., Albert, P., Almahairi,
A., Babaei, Y., Bashlykov, N., Batra, S., Bhargava,
P., Bhosale, S., et al. (2023). Llama 2: Open foun-
dation and fine-tuned chat models. arXiv preprint
arXiv:2307.09288.
Vaswani, A. (2017). Attention is all you need. Advances in
Neural Information Processing Systems.
Xu, H., Kim, Y. J., Sharaf, A., and Awadalla, H. H. (2023).
A paradigm shift in machine translation: Boosting
translation performance of large language models.
arXiv preprint arXiv:2309.11674.
Yuan, L., Chen, D., Chen, Y.-L., Codella, N., Dai, X., Gao,
J., Hu, H., Huang, X., Li, B., Li, C., et al. (2021). Flo-
rence: A new foundation model for computer vision.
arXiv preprint arXiv:2111.11432.
Zhu, J., Xia, Y., Wu, L., He, D., Qin, T., Zhou, W.,
Li, H., and Liu, T.-Y. (2020). Incorporating bert
into neural machine translation. arXiv preprint
arXiv:2002.06823.
ICAART 2025 - 17th International Conference on Agents and Artificial Intelligence
1418