RUNGTA INTERNATIONAL OF JOURNAL OF COMPUTER SCIENCE AND INFORMATION TECHNOLOGY

• Volume 3 (Jan 2026- Dec 2026)
Issue 1

1. Scaling Reasoning in AI: Challenges of Long-Context Understanding in Emerging Models

Nagajayant Nagamani

Author Affiliations

Virtusa, 34 IT Highway, Navallur, Chennai, Tamil Nadu 603103 1/782 Iyyappan Nagar, Iyer Bungalow, Madurai 625014

Technology, Bhilai CG 490024

ABSTRACT:
As artificial intelligence systems handle increasingly complex reasoning tasks, their ability to process extended input sequences remains a critical challenge. This study evaluates the reasoning capabilities of various open-source large language models, including state-space architectures and recurrent frameworks, under varying input lengths. Using a controlled dataset, we investigate performance degradation, reasoning failures, and the impact of structured prompting. Our results reveal key limitations in long-context comprehension, highlighting gaps in information retention and instruction adherence. These findings underscore the need for improved architectures and adaptive learning strategies to enhance multi-hop reasoning in large-scale AI systems.

Keywords: Large Language Models, Long-context Reasoning, State Space Models, RWKV, Mamba, FLenQA, Multi-hop Inference, Attention Mechanisms

2. AI-based Proctoring System for Online Exams

Prabhjeet Kaur1, Aranpreet Singh Saini2

Author Affiliations
1 Department of AI/AIML, Rungta College of engineering and Technology, Bhilai, India-490024
2 Department of CS/DS, Rungta college of engineering and Technology, Bhilai, India-490024

ABSTRACT:
With remote learning and online exams on the rise, maintaining exam integrity is a pressing concern. Manual methods of proctoring in conventional settings are neither scalable nor effective in virtual environments. This project envisages an AI-powered Proctoring System that utilizes computer vision and machine learning algorithms to automate and augment online examination supervision. The system makes use of live face detection, face recognition, and eye-gaze tracking to verify the candidate and observe behavior during the entire length of the exam. It also makes use of object detection and audio analysis for identifying unauthorized devices, multiple faces, or background voices, which can be used as indicators of cheating attempts. The system automatically marks and logs any suspicious activities for examination. Through the integration of deep learning based models like Convolutional Neural Networks (CNNs) for video testing and Natural Language Processing (NLP) for voice anomaly detection, the system provides a secure, scalable, and economical online test environment for a cost-effective alternative to human invigilation. The system does not only reduce the reliance on human invigilators but also increases the fairness and integrity of remote tests.

Keywords: Artificial Intelligence (AI), Online proctoring, Eye gaze tracking, Facial recognition, Automated invigilation.

3. Progressive Learning Strategies for Image Super-Resolution with Deep Neural Network

Prashant Kumar Tamrakar1, Soniya Sinha2, Abhijeet Yadav3, Asha Verma4, Pooja Singh5, Mitali Singh6

Author Affiliations
1 Department of Computer Science & Engineering, Rungta College of Engineering & Technology, Bhilai,
Chhattisgarh, India-490024
2 Department of Computer Science & Engineering, Rungta College of Engineering & Technology, Bhilai,
Chhattisgarh, India-490024
Department of Civil Engineering, Rungta College of Engineering & Technology, Bhilai,
Chhattisgarh, India-490024
4Department of Chemistry, Rungta College of Engineering & Technology, Bhilai,
Chhattisgarh, India-490024
5Department of Civil Engineering, Rungta College of Engineering & Technology, Bhilai,
Chhattisgarh, India-490024
6 Department of Humanities, Rungta College of Engineering & Technology, Bhilai, Chhattisgarh, India-
490024

ABSTRACT:
Image super-resolution (SR) has gained significant attention in recent years due to its potential applications in medical imaging, satellite imagery, and video enhancement. This paper explores progressive learning strategies to enhance the performance of deep neural networks (DNNs) for image super-resolution. By incrementally increasing the complexity of training tasks, our approach allows networks to better generalize across varying image scales and noise levels. We propose a novel framework that integrates multi-stage training, dynamic loss functions, and attention mechanisms to progressively refine image details while preserving contextual integrity. Experimental results demonstrate that the proposed method outperforms state-of- the-art SR models on popular benchmarks, achieving superior performance in terms of both PSNR and SSIM metrics. Furthermore, our approach shows robustness in handling real- world images with diverse degradations, establishing its utility in practical applications.

Keywords: Image Super-Resolution; Progressive Learning; Deep Neural Networks; Attention Mechanisms; PSNR; SSIM; Multi-Stage Training, Image Restoration.

4. A Literature Survey on Deepfake Detection

Richa Sharma1, Tripti Sharma2

Author Affiliations
Rungta College of Engineering and Technology, Bhilai

ABSTRACT:
The quick development of deep learning-based synthetic media production has made deepfake detection an important research topic. False information by deepfakes causes serious threats to Security and privacy. It may alter audio, video, and images to produce incredibly lifelike fakes. There are many deepfake detection methods are examined in this literature review, which are divided into three categories” deep learning, watermarking, and blockchain”. The difficulties with datasets, generalization, and the adversarial robustness of detection models are also covered. Many developments have been done such as multimodal detection frameworks, interpretable explainable AI, and the effect of generative adversarial networks (GANs) on detection effectiveness are highlighted in the study. Lastly, open challenges and future research areas are described to improve deepfake detection’s scalability, performance, and practicality.

Keywords: Deepfake detection, systematic literature review, Deep learning, Watermarking, Blockchain, Dataset

5. Scalable Oversight: Techniques for Aligning Superhuman AI with Human Intent

Bhawana Meher1, Aditya Tiwari2 Priyank Kumar Sahu3, Neha Kumari4, Ambika Singh Sengal4, Ashutosh Tandan5

Author Affiliations
123456 Department of Cyber Security and Data Science, Rungta college of Engineering and Technology, Bhilai, India-490024

ABSTRACT:
As artificial intelligence systems surpass human-level performance in increasingly complex domains, ensuring their alignment with human intent becomes a critical challenge. Scalable oversight refers to methods that enable humans to effectively supervise and guide AI systems, even when those systems operate beyond human comprehension. This paper explores emerging techniques for scalable oversight, including debate, recursive reward modelling, and human-AI collaboration frameworks. We analyze the strengths and limitations of these approaches in maintaining robust alignment while minimizing risks from mis generalization or deceptive behaviors. Additionally, we investigate the role of interpretability tools and automated alignment verification in enhancing oversight scalability. Through case studies in reinforcement learning and large language models, we demonstrate how these techniques can improve AI safety without imposing prohibitive human labour costs. Our findings highlight promising directions for future research, emphasizing the need for adaptive oversight mechanisms that evolve alongside AI capabilities. This work contributes to the broader discourse on AI alignment by providing a systematic evaluation of scalable oversight strategies, offering insights for researchers and practitioners aiming to develop safer, more reliable superhuman AI systems.

Keywords: AI alignment, scalable oversight, superhuman AI, human intent, reinforcement learning, interpretability, AI safety.

RUNGTA INTERNATIONAL OF JOURNAL OF COMPUTER SCIENCE AND INFORMATION TECHNOLOGY