Published as a workshop paper at ICLR 2024

L ARGE L ANGUAGE M ODEL E VALUATION V IA M ULTI

AI AGENTS : P RELIMINARY RESULTS
Zeeshan Rasheed, Muhammad Waseem, Kari Systä & Pekka Abrahamsson
Faculty of Information Technology and Communication Science, Tampere University
Faculty of Information Technology, Jyväskylä University
Tampere & Jyväskylä , Finland
{zeeshan.rasheed & pekka.abrahamsson}@tuni.fi

A BSTRACT
The increasing significance of Large Language Models (LLMs) in both the aca-
demic and industrial sectors can be attributed to their exceptional performance
across a diverse range of applications. As LLMs have become integral to both
research and daily operations, rigorous evaluation is crucial. This assessment is
important not only for individual tasks but also for understanding their societal
impact and potential risks. Despite extensive efforts to examine LLMs from var-
ious perspectives, there is a noticeable lack of multi-agent AI models specifically
designed to evaluate the performance of different LLMs. To address this gap,
we introduce a novel multi-agent AI model that aims to assess and compare the
performance of various LLMs. Our model consists of eight distinct AI agents,
each responsible for retrieving code based on a common description from differ-
ent advanced language models, including GPT-3.5, GPT-3.5 Turbo, GPT-4, GPT-4
Turbo, Google Bard, LLAMA, and Hugging Face. Our developed model utilizes
the API of each language model to retrieve code for a given high-level descrip-
tion. Additionally, we developed a verification agent, tasked with the critical role
of evaluating the code generated by its counterparts. We integrate the HumanEval
benchmark into our verification agent to assess the generated code’s performance,
providing insights into their respective capabilities and efficiencies. Our initial re-
sults indicate that the GPT-3.5 Turbo model’s performance is comparatively better
than the other models. Initially, we provided ten common high-level input descrip-
tions to our proposed model. This preliminary analysis serves as a benchmark,
comparing their performances side by side. Our future goal is to enhance the eval-
uation process by incorporating the Massively Multitask Benchmark for Python
(MBPP) benchmark, which is expected to further refine our assessment. Addition-
ally, we plan to share our developed model with twenty practitioners from various
backgrounds to test our model and collect their feedback for further improvement.

1 I NTRODUCTION
Recent advancements in the field of Large Language Models (LLMs) have brought about signifi-
cant changes in academia and several industries (Hou et al. (2023)). LLMs have become crucial
subjects of study. For example, LLMs such as OpenAI’s GPT Radford et al. (2018), Google’s Bard
Thoppilan et al. (2022), Bing Chat Enterprise Rudolph et al. (2023), LLama Touvron et al. (2023),
and Hugging Face Wolf et al. (2019) demonstrate exceptional skills in understanding, interpreting,
and generating text that closely resembles human communication, proving to be invaluable assets
in various domains. The importance of LLMs extends beyond their technical capabilities to their
exceptional proficiency in processing and analyzing vast amounts of data effectively Rasheed et al.
(2024a). This groundbreaking technology is carving new paths in both research and practical appli-
cations, particularly in the domain of Software Engineering (SE) Wang et al. (2023).
LLMs have recently demonstrated remarkable abilities in processing and interpreting programming
languages, a key aspect of source code modeling Alon et al. (2020), Hellendoorn & Devanbu (2017),
Karampatsis et al. (2020), Rasheed et al. (2024b). These models, driven by sophisticated algorithms
and expansive datasets, have demonstrated an unprecedented ability to comprehend and generate

1
Published as a workshop paper at ICLR 2024

code, offering substantial potential for SE applications Xu et al. (2022), Rasheed et al. (2023).
With the proliferation of various LLMs by leading AI research institutions and corporations, such
as OpenAI’s GPT series, Google’s Bard, DeepMind’s AlphaCode, and others, It becomes crucial
to develop a strong evaluation framework to assess their performance and effectiveness in code
generation tasks.
This paper proposes a novel multi-agent AI model designed for a crucial task. The model utilizes the
strengths of eight distinct AI agents, each interacting with a different LLM to retrieve programming
code based on the same input. This approach aims to capitalize on the unique characteristics and
methodologies of each LLM, offering a comprehensive view of the capabilities in automated code
generation. We use the API key of each model to generate code. The agents work with various LLMs
including GPT-4, GPT-4 Turbo, GPT-3.5, GPT-3.5 Turbo, Google Bard, LLAMA, and Hugging
Face. Our proposed AI agent model facilitates code generation and conducts initial assessments
to guarantee the quality and pertinence of the results. This paper shares early insight into how the
model works in practice.
The crux of our model lies in the eighth agent, the verification agent, whose primary objective is
to perform a meticulous comparison of the code produced by its seven counterparts. This agent
employs HumanEval benchmark Chen et al. (2021) to assesses various attributes of the generated
code, such as syntactic correctness, adherence to the prompt, computational efficiency, and code
accuracy. We utilized pass@k metric into our proposed AI model, enabling a direct evaluation of
the code generated by various models. Pass@k metric is a method in HumanEval benchmark and
a key component of the HumanEval benchmark, serves as a standard measure specifically tailored
for appraising the proficiency of code generation models in creating functional and accurate code
Chen et al. (2022). This benchmark is instrumental in providing a comprehensive and objective
assessment of model performance, ensuring the reliability and effectiveness of our AI model in code
generation tasks.
Our initial result demonstrates that the GPT-3.5 Turbo model exhibits superior performance in gen-
erating accurate code compared to the other models discussed earlier. We tested this by providing
ten common, high-level descriptions as inputs, with each model generating code based on these
descriptions. Remarkably, GPT-3.5 Turbo achieved a 70% accuracy rate in code generation, out-
performing its counterparts. Following closely, GPT-4 Turbo emerged as the second most effective
model, yielding satisfactory results in 6 out of 10 cases. This study significantly contributes to the
ongoing discourse on the practical applications of LLMs and aims to guide stakeholders in making
informed, data-driven decisions when integrating these models into their development workflows.
Our future endeavors are focused on augmenting the evaluation process. We aim to incorporate
the MBPP benchmark Austin et al. (2021) into our multi AI agent model, which we anticipate will
significantly enhance the precision of our evaluations. To this end, we will develop a specialized
AI agent that integrates the MBPP benchmark for a more thorough examination of the generated
code. Furthermore, we intend to expand our input description set from the current 10 to 50, thereby
enabling a more robust and comprehensive analysis of our proposed model. In addition to these
technical enhancements, we are committed to engaging with the wider community by sharing our
model with twenty practitioners from diverse backgrounds. This collaborative approach will not
only provide us with valuable real-world testing but also facilitate the collection of insightful feed-
back, which is essential for the continuous improvement of our model. This final aspect of our plan
underscores our commitment to both technical excellence and practical relevance in the evolving
field of language model evaluation.
Our contribution can be summarized as follow:

• Our study introduces an innovative model for code generation, offering a unique approach
to the evaluation of various LLMs such as GPT-3.5, GPT-3.5 Turbo, GPT-4, GPT-4 Turbo,
Google Bard, LLama, and Hugging Face.

• We integrated the HumanEval benchmark into our verification agent and provided a de-
tailed comparison of each model’s code generation abilities, highlighting their strengths
and weaknesses. Empirical tests with 10 common project descriptions show GPT-3.5 Turbo
leading in accuracy, outperforming others in 7 out of 10 cases.

2
Published as a workshop paper at ICLR 2024

• Our future goal is to enhance the evaluation process by integrating the MBPP benchmark
into our specialized AI agent for comprehensive code examination, and increase input de-
scriptions from 10 to 50 for more detailed analysis. Additionally, we will share the model
with twenty diverse practitioners for real-world testing and feedback, underlining our com-
mitment to technical excellence and practical application in language model evaluation

The rest of the paper is organized as follows: Section 2 reviews related work, and Section 3 describes
the study’s methodology. Primary results are presented in Section 4, and details of our future plans
are outlined in Section 5. The paper concludes in Section 6.

2 R ELATED W ORK

In this section, we briefly present the related work of the study, focusing on existing research. Sec-
tion 2.1 provides an overview of studies concerning Large Language Models (LLMs). Section 2.2
examines works that have developed models for evaluating LLMs.

2.1 L ARGE L ANGUAGE M ODELS

In recent years, LLMs have shown promise in various SE applications Feng et al. (2023). LLMs
are language models consisting of billions of parameters trained from a significant amount of data
and have impressive performance in language processing tasks, including both natural languages
and programming languages Feng et al. (2020), Guo et al. (2022). LLMs designed for processing
and generating human-like text and code, ChatGPT, Google Bard, LLAMA, and Bing Chat En-
terprise are prominent examples within this category Eloundou et al. (2023). These LLM models
are particularly notable for their ability to perform a wide range of language tasks, from translation
and summarizing to question-answering and creative writing, without needing task-specific training.
The core module behind many LLMs such as GPT and BERT are the self-attention module in Trans-
former Vaswani et al. (2017) that serves as the fundamental building block for language modeling
tasks. The introduction of transformer models, as exemplified by OpenAI’s GPT series, marked
a new era in code generation. These models, with their ability to process vast amounts of data
and learn from a wide range of programming languages and styles, significantly outperformed their
predecessors. GPT-3, in particular, demonstrated an unprecedented ability to generate human-like
code, offering not only code completion but also bug fixing, code translation, and even generation
of complex algorithms from natural language descriptions Brown et al. (2020).
Over the past few years, LLMs have demonstrated remarkable efficiency in enabling automatic
programming Xu et al. (2022), Sami et al. (2024). For example, OpenAI’s suite of models has
emerged as a notable force, exhibiting a profound ability in both understanding and generating
code with high accuracy and efficiency Radford et al. (2018). Concurrently, Google’s Bard and
Bing Enterprises have introduced their own iterations of LLMs, tailored to enhance programming
automation with distinct features and methodologies Rudolph et al. (2023). Equally important in
this space are the LLaMA and LLaMA2 models, which have carved out a niche in code generation,
noted for their robustness and adaptability Touvron et al. (2023). Hugging Face’s contributions are
equally significant, offering a diverse array of models that specialize in code generation Wolf et al.
(2019). These models stand out for their user-friendly interfaces and community-driven approach,
making them accessible to a wider range of developers and researchers Chang et al. (2023). Each
of these LLMs has propelled the field of automated code generation forward, presenting a variety of
tools and platforms that cater to different aspects of programming and software development. This
proliferation of models not only highlights the rapid advancements in AI-assisted programming but
also underscores the potential for further innovation and exploration within this domain Guo et al.
(2023). The comparative analysis of these models reveals a landscape rich with possibilities, offering
numerous avenues for research and application in the broader context of artificial intelligence and
machine learning.

2.2 L ARGE L ANGUAGE M ODELS E VALUATIONS

It’s important to assess how well LLMs perform in creating code. This helps make sure these
sophisticated AI models are useful and dependable for actual software development tasks. Checking

3
Published as a workshop paper at ICLR 2024

their performance is essential to confirm that the code they generate is accurate, efficient, and secure,
matching the strict quality demands of the software field. Additionally, regular evaluations help in
identifying areas for improvement, ensuring continuous advancement in AI-driven coding. This
process not only boosts the confidence of developers in using LLMs but also contributes to the
overall growth and innovation in the technology industry. Several researchers, such as Xu et al.
(2022), Chang et al. (2023), and Guo et al. (2023), have conducted surveys of existing LLM models
evaluation for code across various programming languages. The primary goal of these surveys is to
shed more light on the landscape of code modeling design decisions by comparing and contrasting
LLM models.
The development of metrics for evaluating code generated by LLMs has seen significant advance-
ments in recent years, particularly with the introduction of models like CodeBERT Feng et al.
(2020), BertScore Zhang et al. (2019), Codex Chen et al. (2021), PolyCoder Xu et al. (2022), and
the development of new evaluation metrics like ICE-Score Zhuo (2023). CodeBERT developed
by Microsoft, which stands out as a significant milestone in code evaluation. As a bimodal model
trained on both natural language and programming languages, it offers a nuanced approach to assess
the semantic correctness of code snippets Mashhadi & Hemmati (2021). CodeBERT’s ability to
understand the context and functionality of the code has made it a valuable tool for evaluating the
quality of LLM-generated code. After one year. Following CodeBERT, OpenAI introduced Codex,
an AI model based on GPT-3 but fine-tuned for programming tasks. Codex marked a significant step
forward in code generation, capable of understanding and generating complex code sequences. Its
performance necessitated the development of advanced metrics that could effectively evaluate both
the syntactic and logical aspects of the code it generated Chen et al. (2021), Xu et al. (2022) introduce
PolyCoder, an autoregressive language model specifically designed for coding, further advanced the
field. PolyCoder’s specialization in code generation emphasized the need for robust metrics capable
of assessing intricate code structures and the model’s comprehension of programming paradigms.
According to Xu et al. (2022), PolyCoder outperformed codex. Furthermore, using human-written
test suites to evaluate the functional correctness of LLM-generated code can be challenging, espe-
cially in domains where resources are scarce. To address these challenges and limitations, a new
metric, ICE-Score, was proposed. ICE-Score revolutionizes code evaluation by instructing LLMs
for code assessments, offering a novel approach to overcome the obstacles of traditional metrics
Zhuo (2023). It represents a significant step forward in accurately assessing the practical utility and
reliability of code produced by advanced AI models.
However, when it comes to multi-AI agent models, there is a notable gap in the development of
multi-agent systems that can directly retrieve code using an API key and then evaluate the code ac-
curacy of various LLMs. Addressing this gap, we introduce a novel multi-agent AI model designed
to assess and compare the performance of different LLMs. Our work is pioneering in integrating
the pass@k metric into an AI agent for a comparative analysis of code generated by these models.
This approach not only enhances the precision of our evaluation but also provides a more nuanced
understanding of each model’s capabilities. Looking towards the future, we aim to further improve
our evaluation process by incorporating the MBPP benchmark. We believe that this addition will
significantly refine our assessment methods. Furthermore, to ensure the practical applicability and
robustness of our model, we plan to collaborate with twenty practitioners from diverse backgrounds.
By engaging these practitioners in testing our model and collecting their feedback, we seek to gain
valuable insights that will guide further improvements.

3 R ESEARCH M ETHODOLOGY FOR P RELIMINARY A NALYSIS

This research aims to evaluate the performance of various language models through the use of AI
agents. For this purpose, we initially developed a multi-agent model. We utilized an API key for
each LLMs to obtain outputs. By providing a common description, we were able to receive code
generated by different LLMs as outputs. Continuing with this approach, the methodology of our
research is structured into distinct phases. Each phase is intentionally designed to rigorously test
and analyze the capabilities of the LLMs in generating code responses to a common description,
using our multi-agent model system. We formulate the following research questions (RQs):

4
Published as a workshop paper at ICLR 2024

RQ1. How does a multi-agent AI system, with each agent utilizing a different advanced
language model, compare in terms of efficiency and accuracy in generating code from a
common description?

Motivation: The motivation behind this research question stems from the growing importance of
AI in software development and the need to understand the efficacy of different AI approaches.
By exploring how a multi-agent AI system, where each agent employs a distinct advanced lan-
guage model, performs in generating code from a common description, we aim to shed light on
the efficiency and accuracy of these models. This comparison is crucial in an era where the speed
and reliability of automated coding are becoming increasingly vital. It also offers insights into the
strengths and weaknesses of various models, guiding developers and researchers in selecting the
most appropriate AI tools for their specific needs. This research not only contributes to the field of
AI in coding but also has the potential to revolutionize the way we approach software development
and problem-solving in the digital age.

RQ2. What criteria and methods does the verification agent employ to evaluate the perfor-
mance of various language models in code generation tasks?

Motivation: The purpose of this research is to understand how a verification agent checks the work
of different language models when they are used to write code. We want to know what methods and
rules the agent uses to see if the code is good or not. This is important because as we use AI more
for writing code, we need to make sure that the code is correct and works well. By learning how the
agent checks the code, we can make better AI models for coding. This will help people who make
software to use AI in a way that is safe and effective.

3.1 E XPERIMENTAL S ETUP

In this section, we detail our use of API keys from various language models for code generation.
Our primary objective is to assess the accuracy and efficiency of different LLMs. Here, we explain
the process of how these models generate code. In a subsequent section, we will discuss the method-
ology we adopted to evaluate the performance of the generated code. This approach allows for a
comprehensive analysis of the capabilities of different LLMs in coding tasks.
In this paper, we introduce a novel multi-agent AI model consisting of eight distinct AI agents. Each
agent is uniquely configured to interact with a specific advanced language model, including GPT-4,
GPT-4 Turbo, GPT-3.5, GPT-3.5 Turbo, Google Bard, Hugging Face, and LLAMA. The primary
function of these agents is to generate code based on a common description provided to all. This
ensures uniformity in the task and allows for a fair comparison of the output from each language
model. The agents operate simultaneously, retrieving code based on the same set of instructions.
The collected data from each agent is then prepared for subsequent evaluation. Below, we discuss
the entire process of how we utilized each LLM model to generate and evaluate code from given
high-level descriptions. Figure 1 illustrates the entire process of how the API key is used to handle
requests and generate the final code.

3.1.1 O PENAI SERIES

We developed an AI agent model that integrates with multiple versions of the OpenAI GPT models,
including GPT-4, GPT-4 Turbo, GPT-3.5, and GPT-3.5 Turbo. Each AI agent in our model was
configured to interact with a specific version of the OpenAI GPT model. We initiated the process by
providing customized prompts to each agent, embedding a particular project description into these
prompts to maintain consistency across different models. The AI agents, each linked to a distinct
GPT version, were tasked with processing these prompts and generating responses.
Our model efficiently managed the communication with the OpenAI API for each agent, ensuring
that the correct version of the GPT model was utilized for each request. Figure 1 demonstrates
how the API key receives requests and generates responses. The API key provided secure access to
these models, enabling our agents to send and receive data. The responses from the various GPT
versions were then collected and processed to maintain a coherent and continuous dialogue flow.

5
Published as a workshop paper at ICLR 2024

Figure 1: An overview of API request and response processes across different LLMs for Code
generation

By structuring the interaction in several rounds and alternating between the different GPT models,
our model was able to capture a diverse range of AI-generated responses. This setup was pivotal in
assessing the comparative performance and output variations among the different models. The final
outcome of this methodology was a comprehensive collection of responses from each GPT model,
offering a unique opportunity to analyze and understand the nuances in AI-generated content based
on the model version and initial input conditions.

3.1.2 G OOGLE BARD

Google Bard, a sophisticated language model created by Google, played a crucial role in our research
methodology. We employed the Google Bard API to produce descriptive content. Our method
involved using the Bard API alongside Bard Cookies for interaction with Google Bard’s AI system.
This process necessitated authentication via specific cookies, supplied in the form of a dictionary to
the Bard Cookies class. After successful authentication, the system could process and respond to
text-based queries. For example, we input a high-level description into the Bard system, which then
generated code relevant to the input. The responses from Bard offered insightful reflections on the
AI’s ability to comprehend and create intricate, contextually relevant textual content, showcasing its
effectiveness in automated content creation and AI-centric research methodologies.

3.1.3 LLAMA
LLAMA, another prominent language model in our study, was accessed through its designated API
key. This key enabled our algorithm to send high-level descriptions to LLAMA, which then em-
ployed its language understanding capabilities to produce code. We incorporated the use of the
replicate Python library to interact with advanced AI models for code generation. Specifically, we
utilized this library to communicate with the Meta LLAMA-2-70b model. The process involved
sending a high-level prompt to the AI model and model processed the request and generated a cor-
responding output. This output was then iteratively printed, providing us with AI-generated Python
code. This approach demonstrated the model’s capability to understand and generate complex code
based on simple text prompts, showcasing its potential in automating the code generation process
and aiding in software development research.

3.1.4 H UGGING FACE

In our research, we also leveraged Hugging Face’s extensive suite of language models through its
API, focusing specifically on the CodeBERT model. Our method involved providing a high-level
description to the model. Utilizing its advanced natural language processing capabilities, Hugging
Face’s model generated relevant code. This interaction was made possible through the API key,
which served as a conduit, granting our algorithm access to various models under the Hugging Face

6
Published as a workshop paper at ICLR 2024

platform. Each model contributed distinct strengths to the code generation process, demonstrating
the versatility and efficacy of Hugging Face’s AI solutions in our research.

3.2 I NPUT OF C ODE E VALUATION

In this research, we focused on the development of an AI agent model capable of robustly interpret-
ing natural language prompts and generating corresponding code based on the provided descriptions.
Our methodology involved utilizing common natural language descriptions as inputs, which were
then processed by various state-of-the-art LLMs to generate code outputs.
To effectively evaluate the performance of these LLMs in code generation, we curated a diverse set
of 10 descriptive prompts, each originating from different project backgrounds. This diversity in
input descriptions was deliberately chosen to encompass a wide range of scenarios and complexities
that developers might encounter in real-world applications. The specifics of these input descriptions
are presented below.
The box below displays the input prompts used for evaluating various LLMs, thereby ensuring a
comprehensive assessment of the LLMs’ code generation capabilities. By encompassing a wide
array of subjects and complexities in our input descriptions, our approach aims to rigorously test
the adaptability and accuracy of the LLMs in converting natural language instructions into func-
tional code across different contexts. This methodological approach provides a robust framework
for assessing the efficacy of LLMs in understanding and translating human language into executable
code.

01. A simple Python program for creating and using flashcards for study, allowing users to
test their knowledge on various subjects
02. An interactive Python game that tests users on various academic subjects and provides
instant feedback
03. A short, interactive, text-based adventure game written in Python, where players make
choices that influence the outcome of the story.
04. A Python-based Pomodoro timer that helps users apply the Pomodoro technique (25
minutes of work followed by a 5-minute break) to boost productivity.
.
.
.
.
10. A Python-based basic arithmetic calculator capable of performing operations like addi-
tion, subtraction, multiplication, and division

3.3 M ETHODOLOGY FOR C ODE E VALUATION

The evaluation process involves analyzing various aspects of the code generated by each language
model. Key metrics include the accuracy of the code in relation to the given description, efficiency in
terms of execution and resource usage, and the innovation demonstrated in the solutions. This study
employs a structured approach to assess the code generation capabilities of various advanced LLMs,
including GPT-4, GPT-4 Turbo, GPT-3.5, GPT-3.5 Turbo, LLAMA, Hugging Face, and Google
BARD. We developed a specialized AI agent with the primary objective of evaluating the accuracy
and reliability of code produced by these models. The cornerstone of our evaluation methodology is
the integration of the HumanEval benchmark within our AI agent.
This model utilized the HumanEval benchmark, a widely recognized standard for assessing code
quality and effectiveness. For evaluation, we employed the pass@k metric, a rigorous method to
determine the viability of generated code. Our model receives a description as input, and in response,
it obtains code from various LLMs. We provided ten distinct descriptions, each generating a set
of potential solutions. Using the pass@k metric, we evaluated these solutions to determine their
correctness. Mathematically, pass@k can be expressed as the probability that at least one of the top
k solutions generated by the model is correct.

7
Published as a workshop paper at ICLR 2024

"
#
n−c
k

pass@k := EProblems 1 − n
k

If one solution out of the k attempts is found correct, we mark the instance as successful and proceed
to the next description. In this way, we systematically assess the performance of LLMs across
multiple coding tasks, ensuring a comprehensive analysis of their capabilities in code generation.
The methodology’s implementation involves iterating over the set of descriptions, generating code
for each, and applying the pass@k metric to assess each set of solutions. This approach allows for
an empirical evaluation of LLMs in terms of both accuracy and efficiency in coding tasks.

4 P RELIMINARY R ESULTS

In this section, we present the initial results of the proposed multi agents model for evaluation of
various LLM-based model. The primarily results are shown in Table 1. So far our findings indicate
that the GPT 3.5 Turbo model perform relatively better than others LLM based model for generating
accurate code by given high level description. Below, we present the results of our proposed model
in Section 4.1, specifically reporting the outcomes of RQ1. Additionally, we also provide the detail
criteria and methods that we employed in verification agent to evaluate the performance of various
language models in Section 4.2.

4.1 P ROPOSED M ULTI AGENT M ODEL (RQ1)

In this section of our research, we conducted an evaluation of code performance across various
language models. This assessment was carried out by developing a sophisticated algorithm that
utilized the API keys of the respective language models. The core objective was to methodically
analyze and compare the output quality and efficiency of each model when tasked with generating
code from high-level descriptions. In the pursuit of understanding the capabilities of various LLMs
in the context of code generation, we conducted a comprehensive evaluation using the HumanEval
benchmark to gauge the accuracy of the code produced by each model. Table 1 presents a synthe-
sized view of this evaluation, showcasing the performance of different LLMs across identical input
descriptions.
Upon early testing and analysis, our results indicated that GPT-3.5 Turbo exhibited superior perfor-
mance compared to other language models like GPT-4, GPT-4 Turbo, Google Bard, Hugging Face,
and LLAMA. Specifically, GPT 3.5 Turbo provided accurate results for seven out of the ten inputs.
This achievement marks a significant margin of excellence compared to its counterparts. In con-
trast, other LLM models in the study produced fewer accurate outputs, as can be observed from the
comparative data in the Table 1.
Table 1 enumerates seven different LLMs, each produced by renowned organizations with varying
parameter sizes—from 355M by Hugging Face to 1.96 trillion by GPT-4 Turbo. Each model was
provided with the same set of ten input descriptions to process and generate code. The primary
focus of the evaluation was on the accuracy of the results, which refers to the functional correctness
of the generated code as per the given description. Additionally, a quality rating, depicted with stars,
provides a subjective assessment of the code based on criteria such as readability, efficiency, and
adherence to best practices.
The data shows a varied performance across models. GPT-3.5 Turbo, with 154 billion parameters,
leads with the highest number of accurate results, successfully generating correct code for seven
out of ten descriptions. This is followed closely by GPT-4 Turbo, which each returned six accu-
rate pieces of code. Notably, despite Google Bard’s significantly larger parameter count, it did not
outperform GPT-3.5 Turbo, indicating that larger model size does not necessarily equate to better
performance in specific tasks such as code generation.
The results underscore the efficacy of GPT-3.5 Turbo in this domain, with a strong blend of accuracy
and high quality, as reflected by its four-star rating. The other models, while demonstrating varying
levels of proficiency, highlight the competitive landscape of LLMs and their potential for software
engineering applications.

8
Published as a workshop paper at ICLR 2024

This analysis provides initial insights into the current state of LLMs in code generation tasks, under-
scoring the importance of specialized benchmarks like HumanEval in evaluating AI agents. It also
sets the stage for further discussions on model optimization, the impact of parameter scale, and the
potential for LLMs to revolutionize the field of automated code generation.

4.2 I NITIAL E VALUATION F RAMEWORK (RQ2)

In the Initial Evaluation Framework section, our focus shifted towards the validation and verification
of the code generated by these models. In our study, we deployed an AI multi-agent model to
evaluate the code generation capabilities of various LLMs, including GPT-4, GPT-4 Turbo, GPT-
3.5, GPT-3.5 Turbo, Google Bard, LLaMA, and Hugging Face. The performance assessment of
these models was based on their ability to generate accurate and functional code in response to a set
of predefined descriptions.
Our evaluation strategy involved presenting each LLM with 10 unique code generation tasks. The
performance of each model was assessed using the HumanEval benchmark, which is a widely rec-
ognized method for evaluating code synthesis. In this project, we developed multi AI agent model
to generate and evaluate the generated code. The task of one agent was to employ the HumanEval
benchmark and autonomously test the code generation capabilities of AI models Each task in Hu-
manEval requires the AI to write a function that satisfies a given specification, and the models’
responses are evaluated based on their correctness and efficiency.
One of the key metrics used in our assessment was the pass@k metric, a standard measure for eval-
uating code generation models. The pass@k metric evaluates whether a correct solution is found
within k attempts, where k is a predefined number of generated code samples. This metric is partic-
ularly useful in scenarios where multiple solutions are possible, as it accounts for the probabilistic
nature of code generation by LLMs. In our experiment, we utilized a pass@1 approach, meaning
that we evaluated the models based on their first attempt at generating a solution. This approach was
chosen to reflect a more realistic scenario where a developer would use the model’s first output.
The results were revealing: GPT-3.5 Turbo outperformed other models in our tests, delivering correct
and efficient code solutions in 7 out of 10 tasks. This indicates a 70% success rate on the first attempt,
which is a significant achievement considering the complexity of the tasks involved. In comparison,
other models like GPT-4 and GPT-4 Turbo showed varying levels of performance, but none matched
the consistency of GPT-3.5 Turbo.
This evaluation highlights the effectiveness of GPT-3.5 Turbo in code generation tasks, particularly
in its ability to understand and translate complex task descriptions into functional code. The use of
HumanEval benchmark and the ‘pass@k‘ metric provided a robust framework for this assessment,
allowing for a fair and comprehensive comparison across different LLMs.

Table 1: Workflow of the proposed model: Application across diverse datasets and output formats
LLM Model Product Parameter I/P Description Accurate Result Quality
GPT-4 Turbo OpenAI 1.96 trillion 10 6 ★★★✩✩
GPT-4 OpenAI 1.76 trillion 10 5 ★★★✩✩
GPT-3.5 Turbo OpenAI 154 billion 10 7 ★★★★✩
GPT-3.5 OpenAI 125 billion 10 5 ★★★✩✩
Google Bard Google 1.56 trillion 10 4 ★★✩✩✩
LLama Meta 70 billion 10 2 ★✩✩✩✩
Hugging Face Hugging Face 355M 10 2 ★✩✩✩✩

5 F UTURE W ORK

As we work to improve how we evaluate LLMs for code generation, our current methodology em-
ploys the HumanEval benchmark within our AI agent framework. HumanEval has served as a robust
starting point, offering a suite of programming problems that models must solve, thereby providing
a quantitative measure of their code generation abilities. However, to broaden our evaluation’s scope
and depth, we are preparing to incorporate the MBPP. The MBPP benchmark, with its larger and

9
Published as a workshop paper at ICLR 2024

more diverse set of programming tasks, will enable a more extensive assessment of LLMs’ capabil-
ities.
In conjunction with the adoption of MBPP, we recognize the indispensable value of human judgment
in model evaluation. To this end, we have embarked on a collaboration with twenty practitioners
from a spectrum of backgrounds. These industry experts, academics, and developers will bring their
unique perspectives to bear on the task of manually evaluating our proposed model’s performance.
Their insights will not only validate the findings from the MBPP benchmark but also provide nu-
anced understanding of the model’s practical effectiveness and areas for improvement.
This dual approach, combining automated benchmarks with human expertise, wil offer a compre-
hensive evaluation framework. It will ensure that the strengths and weaknesses of LLMs in code
generation are fully understood, paving the way for targeted enhancements. With this initiative, we
aim to set a new standard for the evaluation of LLMs, one that balances the scalability of automated
testing with the depth of human review.

6 C ONCLUSIONS

In this paper, we introduced a novel multi-agent AI model designed to assess and compare the
performance of various LLMs. Our system comprises eight distinct AI agents, each tasked with the
retrieval of code based on a shared high-level description from several advanced language models,
including GPT-3.5, GPT-3.5 Turbo, GPT-4, GPT-4 Turbo, Google Bard, LLAMA, and Hugging
Face. Utilizing the APIs provided by these models, our agents can retrieve code that is then evaluated
by a specialized verification agent. This verification agent employs the HumanEval benchmark to
measure performance, offering valuable insights into the strengths and efficiencies of each model.
Initial findings have highlighted GPT-3.5 Turbo’s superior performance relative to its counterparts.
As we look to the future, our ambition is to further enhance the evaluation framework. The in-
tegration of the MBPP represents a significant next step in this process, promising to sharpen our
assessment tools and provide even more granular insights. Moreover, we are committed to engaging
with the broader community by involving twenty practitioners from a variety of backgrounds to in-
teract with and evaluate our model. Their feedback will be indispensable, providing a practical edge
to our theoretical findings and helping to guide the model’s continuous improvement.
Our work lays the groundwork for a dynamic and robust evaluation platform that not only bench-
marks the current state of LLMs but also adapts and evolves with the field. The insights gained
through our research will serve as a cornerstone for future advancements in AI-driven code gen-
eration, facilitating developments that are as practical as they are innovative. Through rigorous
assessment and community engagement, we aspire to drive the field forward, setting new standards
for performance and applicability in the ever-growing expanse of language model technology.

7 ACKNOWLEDGMENT

We express our sincere gratitude to Business Finland for their generous support and funding of
our project. Their commitment to fostering innovation and supporting research initiatives has been
instrumental in the success of our work.

R EFERENCES
Uri Alon, Roy Sadaka, Omer Levy, and Eran Yahav. Structural language models of code. In Inter-
national conference on machine learning, pp. 245–256. PMLR, 2020.

Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan,
Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, et al. Program synthesis with large language
models. arXiv preprint arXiv:2108.07732, 2021.

Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal,
Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are
few-shot learners. Advances in neural information processing systems, 33:1877–1901, 2020.

10
Published as a workshop paper at ICLR 2024

Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan
Yi, Cunxiang Wang, Yidong Wang, et al. A survey on evaluation of large language models. ACM
Transactions on Intelligent Systems and Technology, 2023.

Bei Chen, Fengji Zhang, Anh Nguyen, Daoguang Zan, Zeqi Lin, Jian-Guang Lou, and Weizhu
Chen. Codet: Code generation with generated tests. arXiv preprint arXiv:2207.10397, 2022.

Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared
Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large
language models trained on code. arXiv preprint arXiv:2107.03374, 2021.

Tyna Eloundou, Sam Manning, Pamela Mishkin, and Daniel Rock. Gpts are gpts: An early look
at the labor market impact potential of large language models. arXiv preprint arXiv:2303.10130,
2023.

Yunhe Feng, Sreecharan Vanam, Manasa Cherukupally, Weijian Zheng, Meikang Qiu, and Haihua
Chen. Investigating code generation performance of chat-gpt with crowdsourcing social data. In
Proceedings of the 47th IEEE Computer Software and Applications Conference, pp. 1–10, 2023.

Zhangyin Feng, Daya Guo, Duyu Tang, Nan Duan, Xiaocheng Feng, Ming Gong, Linjun Shou, Bing
Qin, Ting Liu, Daxin Jiang, et al. Codebert: A pre-trained model for programming and natural
languages. arXiv preprint arXiv:2002.08155, 2020.

Daya Guo, Shuai Lu, Nan Duan, Yanlin Wang, Ming Zhou, and Jian Yin. Unixcoder: Unified
cross-modal pre-training for code representation. arXiv preprint arXiv:2203.03850, 2022.

Zishan Guo, Renren Jin, Chuang Liu, Yufei Huang, Dan Shi, Linhao Yu, Yan Liu, Jiaxuan Li,
Bojian Xiong, Deyi Xiong, et al. Evaluating large language models: A comprehensive survey.
arXiv preprint arXiv:2310.19736, 2023.

Vincent J Hellendoorn and Premkumar Devanbu. Are deep neural networks the best choice for
modeling source code? In Proceedings of the 2017 11th Joint meeting on foundations of software
engineering, pp. 763–773, 2017.

Xinyi Hou, Yanjie Zhao, Yue Liu, Zhou Yang, Kailong Wang, Li Li, Xiapu Luo, David Lo, John
Grundy, and Haoyu Wang. Large language models for software engineering: A systematic litera-
ture review. arXiv preprint arXiv:2308.10620, 2023.

Rafael-Michael Karampatsis, Hlib Babii, Romain Robbes, Charles Sutton, and Andrea Janes.
Big code!= big vocabulary: Open-vocabulary models for source code. In Proceedings of the
ACM/IEEE 42nd International Conference on Software Engineering, pp. 1073–1085, 2020.

Ehsan Mashhadi and Hadi Hemmati. Applying codebert for automated program repair of java simple
bugs. In 2021 IEEE/ACM 18th International Conference on Mining Software Repositories (MSR),
pp. 505–509. IEEE, 2021.

Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language under-
standing by generative pre-training. 2018.

Zeeshan Rasheed, Muhammad Waseem, Kai-Kristian Kemell, Wang Xiaofeng, Anh Nguyen Duc,
Kari Systä, and Pekka Abrahamsson. Autonomous agents in software development: A vision
paper. arXiv preprint arXiv:2311.18440, 2023.

Zeeshan Rasheed, Muhammad Waseem, Aakash Ahmad, Kai-Kristian Kemell, Wang Xiaofeng,
Anh Nguyen Duc, and Pekka Abrahamsson. Can large language models serve as data analysts?
a multi-agent assisted approach for qualitative data analysis. arXiv preprint arXiv:2402.01386,
2024a.

Zeeshan Rasheed, Muhammad Waseem, Mika Saari, Kari Systä, and Pekka Abrahamsson. Code-
pori: Large scale model for autonomous software development by using multi-agents. arXiv
preprint arXiv:2402.01411, 2024b.

11
Published as a workshop paper at ICLR 2024

Jürgen Rudolph, Shannon Tan, and Samson Tan. War of the chatbots: Bard, bing chat, chatgpt, ernie
and beyond. the new ai gold rush and its impact on higher education. Journal of Applied Learning
and Teaching, 6(1), 2023.
Abdul Malik Sami, Zeeshan Rasheed, Kai-Kristian Kemell, Muhammad Waseem, Terhi Kilamo,
Mika Saari, Anh Nguyen Duc, Kari Systä, and Pekka Abrahamsson. System for systematic lit-
erature review using multiple ai agents: Concept and an empirical evaluation. arXiv preprint
arXiv:2403.08399, 2024.
Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze
Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, et al. Lamda: Language models for dialog
applications. arXiv preprint arXiv:2201.08239, 2022.
Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée
Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and
efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023.
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez,
Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural informa-
tion processing systems, 30, 2017.
Junjie Wang, Yuchao Huang, Chunyang Chen, Zhe Liu, Song Wang, and Qing Wang. Software test-
ing with large language model: Survey, landscape, and vision. arXiv preprint arXiv:2307.07221,
2023.
Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi,
Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, et al. Huggingface’s transformers:
State-of-the-art natural language processing. arXiv preprint arXiv:1910.03771, 2019.
Frank F Xu, Uri Alon, Graham Neubig, and Vincent Josua Hellendoorn. A systematic evaluation of
large language models of code. In Proceedings of the 6th ACM SIGPLAN International Sympo-
sium on Machine Programming, pp. 1–10, 2022.
Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluat-
ing text generation with bert. arXiv preprint arXiv:1904.09675, 2019.
Terry Yue Zhuo. Large language models are state-of-the-art evaluators of code generation. arXiv
preprint arXiv:2304.14317, 2023.

12