Volume 10, Issue 3, March – 2025 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25mar1924

AI-Powered Exam Assessment System for

Handwritten Answer Sheets
Naman Agnihotri ; Harshvardhan Grandhi; Dhanashri Patil ; Sanika Kharade
Publication Date: 2025/04/16
Abstract: This paper introduces an AI-powered exam assessment system designed to automate the evaluation of handwritten
answer sheets, encompassing both textual answers and diagrams. The system addresses the inherent limitations of traditional
manual grading methods, such as their labor-intensive nature, susceptibility to human error, and time consumption. In
contrast to conventional Optical Character Recognition (OCR) solutions that struggle with handwriting diversity and visual
content, the proposed system directly interprets both text and visual data, enabling accurate and efficient grading of diverse
student responses. By leveraging AI models with multimodal capabilities, the system effectively compares student answers
with predefined question papers and answer keys to ensure objective and consistent grading. This innovative approach offers
a scalable and cost-effective solution for educational institutions, significantly reducing the time and resources required for
manual evaluations while enhancing the accuracy and fairness of the assessment process.

Keywords: Large Language Models (LLMs), Vision Language Models (VLMs), Handwritten Answer Assessment, Automated Grading,
AI in Education, Multimodal Assessment, Diagram Evaluation, Scalable Assessment System.

How to Cite: Naman Agnihotri ; Harshvardhan Grandhi; Dhanashri Patil ; Sanika Kharade (2025) AI-Powered Exam Assessment
System for Handwritten Answer Sheets. International Journal of Innovative Science and Research Technology, 10(3), 3094-3097.
https://doi.org/10.38124/ijisrt/25mar1924

I. INTRODUCTION capacity to interpret visual data, a novel AI-powered system

can be constructed to automate the evaluation of both textual
 Challenges of Manual Grading responses and graphical representations within exam papers,
Manual exam paper grading poses substantial challenges thereby establishing a foundation for accurate, scalable, and
for educational institutions, characterized by its labor- cost-effective assessments, aiming to revolutionize the grading
intensive and time-consuming nature, susceptibility to human process by circumventing the necessity for OCR and
error resulting in inconsistent evaluations, and amplification capitalizing on the advanced capabilities of AI, ultimately
by rising student numbers and complex exam formats addressing the inefficiencies inherent in traditional manual
involving written responses and detailed diagrams; grading while ensuring consistent and objective evaluation
consequently, this process leads to delayed feedback and outcomes.
elevated operational costs.
II. METHODS
 Limitations of Existing Automated Solutions
Traditional automated grading solutions, relying heavily  System Overview
on Optical Character Recognition (OCR), face significant The AI-Powered Exam Assessment system automates the
limitations due to OCR's struggles with diverse handwriting evaluation of handwritten answer sheets by integrating LLMs
styles and its inability to evaluate visual elements such as and VLMs to process and interpret student responses.
diagrams, thereby underscoring the necessity for a more
sophisticated, AI-driven approach to exam grading.  Handwritten Text Interpretation
To ensure accurate evaluation, the system employs a
 Proposed AI-Powered Solution precise mapping mechanism that correlates questions from the
The convergence of Large Language Models (LLMs) answer sheets to their respective answers, thereby
and Vision Language Models (VLMs) offers a distinctive guaranteeing that the AI models assess the correct responses;
opportunity to transcend the limitations of conventional furthermore, Large Language Models (LLMs) are integrated
automated grading methods, which are often constrained by to interpret and evaluate handwritten text, accommodating a
Optical Character Recognition (OCR) inadequacies; by broad spectrum of handwriting styles to achieve reliable
harnessing LLMs' ability to comprehend and process textual comprehension of textual answers.
content with enhanced precision, coupled with VLMs'

IJISRT25MAR1924 www.ijisrt.com 3094

Volume 10, Issue 3, March – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25mar1924

III. RESULTS
 Diagram and Visual Content Assessment
VLMs are utilized to assess diagrams and other visual  System Capabilities
elements present in the answer sheets, enabling a more The AI-Powered Exam Assessment system has
comprehensive evaluation than traditional text-based systems. demonstrated the capability to automate the grading of
handwritten answer sheets, accurately evaluating both textual
 Automated Grading Process and visual content.
The AI models compare student responses to a
predefined question paper and answer key, automating the  Accuracy and Effectiveness
grading process with consistent and objective criteria. Through rigorous training on an extensive dataset of
answer sheets, the system has attained a high level of precision
 Development Approach in aligning questions with their corresponding answers,
Employing a modular design, the system is engineered to thereby ensuring accurate evaluation; furthermore, it has
facilitate scalability and accommodate future enhancements, demonstrated notable efficacy in processing a wide array of
thereby ensuring its adaptability to increasing demands and handwriting styles and adeptly evaluating diverse diagram
technological advancements; moreover, an agile development types, showcasing its robustness and adaptability to varied
methodology is adopted to promote iterative progress and exam formats.
responsiveness to evolving requirements, enabling continuous
improvement and alignment with dynamic educational needs.  Validation through Testing

Fig 1. Model's Assessment v/s Actual Assessment

IJISRT25MAR1924 www.ijisrt.com 3095

Volume 10, Issue 3, March – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25mar1924

Fig 2. Model’s Deviation from Actual Assessment

The system's performance has been validated through  Addressing Inefficiencies

rigorous testing procedures, including load testing, stress By automating the intricate evaluation of handwritten
testing, scalability testing, endurance testing, volume testing, answer sheets, this system effectively confronts and mitigates
concurrency testing, and software accuracy testing. the inherent inefficiencies and limitations that have long
plagued manual grading, such as time constraints, human
Table 1. Comparison of Character Error Rate (CER) for error, and inconsistent application of grading criteria, thereby
Different OCR Engines on the DUDE Benchmark streamlining the entire assessment workflow.
OCR Engine Category CER (DUDE
 Comprehensive and Objective Evaluation
benchmark)
The system's unique ability to accurately and
comprehensively assess both textual answers and complex
Our Solution Multimodal 15.14%
diagrams ensures a more holistic and objective evaluation of
LLM‑based
student work, surpassing the limitations of traditional methods
Open‑source 38.94% that often struggle with nuanced interpretations and visual
Tesseract OCR
representations, ultimately leading to a more accurate
EasyOCR Open‑source 64.41% reflection of student understanding.

PaddleOCR Open‑source 36.40%  Scalability and Cost-Effectiveness

The system's inherent scalability and operational
MMOCR Open‑source 60.67% efficiency render it exceptionally well-suited for handling the
ever-increasing volumes of exam data, thereby offering a cost-
Our solution has the least amount of error rate when it comes effective and sustainable solution for educational institutions
to handwritten text. seeking to optimize their assessment processes without
compromising on accuracy or thoroughness.
IV. DISCUSSION
 Future Enhancements
 Transformation of Grading Processes While the current model has demonstrated remarkably
The outcomes derived from this project unequivocally promising results, ongoing and dedicated efforts to
highlight the profound transformative potential inherent in AI- meticulously fine-tune its parameters and strategically expand
powered systems, signaling a paradigm shift in the way its capabilities will further enhance its accuracy, versatility,
traditional exam grading processes are conducted, moving and adaptability to diverse exam formats, ensuring that it
from a labor-intensive, subjective exercise to an automated, remains at the forefront of AI-driven assessment technologies.
objective, and efficient methodology

IJISRT25MAR1924 www.ijisrt.com 3096

Volume 10, Issue 3, March – 2025 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165 https://doi.org/10.38124/ijisrt/25mar1924

V. CONCLUSION understanding among educators and students, thereby

demystifying the technology and promoting its responsible
 Advancement in Exam Evaluation adoption.
The development and implementation of the AI-Powered
Exam Assessment system signify a substantial and  Accountability
groundbreaking advancement in the automation of evaluating To ensure responsible and reliable AI-driven evaluations,
handwritten answer sheets, marking a pivotal shift towards clear lines of accountability must be established; furthermore,
more sophisticated and efficient assessment methodologies. integrating human oversight and review mechanisms into the
system is essential to address any errors or discrepancies that
 Benefits over Traditional Methods may arise in automated grading, thereby maintaining accuracy
By strategically leveraging the advanced capabilities of and trustworthiness in the assessment process.
Large Language Models (LLMs) and Vision Language
Models (VLMs), the system provides a significantly faster, REFERENCES
demonstrably more accurate, and inherently scalable solution
when juxtaposed with traditional manual grading methods, [1]. Evaluating Students’ Descriptive Answers Using Natural
which are often characterized by their time-consuming nature Language Processing and Artificial Neural Networks
and susceptibility to human error. (IJCRT), Dec 2017
[2]. Evaluation of Descriptive Answer Sheet Using Artificial
 Ensuring Thorough and Equitable Assessment Intelligence (IJESRT), April 2019
The system's unique capacity to meticulously assess both [3]. AI-based Test Automation: A Grey Literature Analysis
textual answers and intricate diagrams ensures a more ,May 2021
comprehensive, thorough, and fundamentally equitable [4]. Automation of Answer Scripts Evaluation - A Review,
evaluation process, mitigating the inconsistencies and biases Mar 2021
that can arise from purely human-driven assessments. [5]. AI-Based Automatic Subjective Answer Evaluation
System (ZKG International), April 2024
 Impact on Educational Institutions [6]. Automated Essay Scoring Using Machine Learning
The implementation of this system ultimately yields Techniques (Journal of Educational Measurement), 2020.
substantial benefits for educational institutions by fostering a [7]. Deep Learning for Diagram Understanding in
culture of fairness, ensuring consistency in grading standards, Educational Assessments (IEEE Transactions on Pattern
and dramatically enhancing the overall efficiency of exam Analysis and Machine Intelligence), 2022.
grading processes, thereby optimizing resource allocation and [8]. A Hybrid Approach for Handwritten Text Recognition in
improving educational outcomes. Examination Scripts (Pattern Recognition Letters), 2019.
[9]. Large Language Models for Educational Assessment:
VI. ETHICAL CONSIDERATIONS Opportunities and Challenges (Educational Technology
Research and Development), 2023.
 Data Privacy and Security [10]. Vision Transformers for Visual Question Answering on
The paramount importance of ensuring secure storage Educational Diagrams (Conference on Computer Vision
and handling of student data, encompassing answer sheets and and Pattern Recognition), 2023.
evaluation results, necessitates the implementation of robust [11]. Developing a Scalable AI-Based Grading System for
measures to safeguard sensitive information from Complex Exam Formats (International Journal of
unauthorized access and potential breaches, thereby upholding Artificial Intelligence in Education), 2022.
student privacy and institutional integrity. [12]. The Impact of AI-Driven Automated Grading on Teacher
Workload and Student Feedback (Computers &
 Fairness and Bias Education), 2021.
To mitigate biases and ensure fairness in evaluations, AI [13]. Ethical Considerations in AI-Based Automated
models must be trained on diverse and representative datasets Assessment (Journal of Educational Technology &
that accurately reflect the student population; furthermore, Society), 2020.
continuous monitoring and refinement of these models are [14]. A Survey of Automated Short Answer Grading
essential to proactively identify and address any potential Techniques (Natural Language Engineering), 2018.
sources of bias, thereby upholding equitable assessment [15]. Evaluating the Reliability and Validity of AI-Based
practices. Diagram Assessment Tools (Applied Measurement in
Education), 2023.
 Transparency and Explainability
Although AI models offer efficient evaluations, it
remains crucial to foster transparency in the grading process;
specifically, providing detailed feedback and explanations for
AI-driven assessments can significantly enhance trust and

IJISRT25MAR1924 www.ijisrt.com 3097