Veritatis Gaudium: Multidisciplinary Journal of Postgraduate Studies

Since the advent of GPT-0, the development of generative artificial intelligence for text generation has undergone rapid and transformative progress over the past decade. There has been a growing demand for Artificial Intelligence tools. Several Artificial Intelligence models have been released for tasks such as text generation. In this study, we conducted a comprehensive evaluation of leading generative Artificial Intelligence models for text generation. Our study explored the various architectural nuances, training methodologies, strengths, limitations, and application domains of these models. The models studied are: Recurrent Neural Networks (RNNs), Transformers, GPT2/3, BART, FLAN T5 and their derivatives. A total of 35 research works were studied to establish a robust theoretical foundation and to ascertain the current state of the field. Results obtained from the study showed that generative AI models have matured significantly over the past decade. However, despite the enormous growth, they exhibit varying degrees of strength and limitations that still persist. This, however, depends on the design objectives of the models, architecture, and training standard. Furthermore, the study showed that no single model outperformed other models on all metrics. We also observed that models like InstructGPT and ChatGPT demonstrate how fine-tuning with human feedback and reinforcement mechanisms can lead to better controllability, ethical behaviour, and adaptability. Further results showed that FLAN T5 showcases the effectiveness of instruction-based multitask training in improving generalization across several fields. Meanwhile, models such as LlaMA 2 and Gemma illustrate the feasibility of deploying efficient, albeit less sophisticated, generative models in edge environments.

Open Full Text (PDF)