IITM LOGO V17 FILE 1

IITM Journal of Information Technology

ISSN (P) 2395-5457 | Single Blind Peer Reviewed Journal

Published By
INSTITUTE OF INNOVATION IN TECHNOLOGY & MANAGEMENT
Affiliated to GGSIPU, NAAC Grade ‘A’, ISO 14001:2015, 17020:2012, 21001:2018 & 50001:2018 Certified,
A Grade by GNCTD, A++ Grade by SFRC

Large Language Models and Agentic AI: How Modern AI Systems Are Evolving Into Autonomous Agents

Deepti Sharan1, Nehuti2
 1Department of AI-ML: Business Application, McCombs School of Business, UT Austin
 2Department of Computer Science, Bharati Vidyapeeth College of Engineering, New Delhi, India 

Abstract:  Large Language Models (LLMs) have undergone a dramatic transformation from static text-completion engines to dynamic, tool-wielding autonomous agents capable of planning, reasoning, and executing multi-step tasks in real-world environments. This paper provides a comprehensive survey of the architectural principles, training paradigms, and emergent capabilities that underpin modern LLMs such as GPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro. We examine the progression from transformer-based language models toward agentic AI systems that integrate memory modules, external tools, and multi-agent coordination frameworks. Key topics include the Retrieval- Augmented Generation (RAG) paradigm, chain-of- thought and tree-of thought prompting strategies, the ReAct framework, and multi-agent orchestration platforms such as AutoGen and CrewAI. We further present a structured experimental comparison of Chain-of-Thought (CoT) and ReAct prompting paradigms across 120 standardized tasks spanning multi-step question answering, logical inference, and knowledge-retrieval scenarios, provide in reproducible methodology and statistical analysis. We also discuss open challenges in safety, alignment, hallucination mitigation, and computational costs associated with large-scale deployment. Our analysis reveals that while agentic AI demonstrates remarkable potential in software engineering, scientific research, and enterprise automation, significant hurdles in reliability, explainability, and ethical governance must be addressed before wide- scale deployment can be responsibly achieved.

Keywords:  Large Language Models, Agentic AI, Autonomous Agents, Transformer Architecture, Chain- of-Thought, Retrieval Augmented Generation, Multi- Agent Systems, AI Safety.

References:

  1. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, "Attention is all you need," in Advances in Neural Information Processing Systems (NeurIPS), vol. 30, 2017.
  2. T. Brown, B. Mann, N. Ryder et al., "Language models are few-shot learners," in NeurIPS, vol. 33, pp. 1877-1901, 2020.
  3. J. Kaplan, S. McCandlish, T. Henighan et al., "Scaling laws for neural language models," arXiv preprint arXiv:2001.08361, 2020.
  4. J. Hoffmann, S. Borgeaud, A. Mensch et al., "Training compute-optimal large language models," in NeurIPS, 2022.
  5. J. Wei, X. Wang, D. Schuurmans et al., "Chain-ofthought prompting elicits reasoning in large language models," in NeurIPS, vol. 35, 2022.
  6. J. Wei, Y. Tay, R. Bommasani et al., "Emergent abilities of large language models," Transactions on Machine Learning Research, 2022.
  7. S. Yao, J. Zhao, D. Yu et al., "ReAct: Synergizing reasoning and acting in language models," in ICLR, 2023.
  8. S. Yao, D. Yu, J. Zhao et al., "Tree of thoughts: Deliberate problem solving with large language models," in NeurIPS, 2023.
  9. L. Ouyang, J. Wu, X. Jiang et al., "Training language models to follow instructions with human feedback," in NeurIPS, vol. 35, 2022.
  10. Anthropic, "Constitutional AI: Harmlessness from Al feedback," arXiv preprint arXiv:2212.08073, 2022.
  11. P. Lewis, E. Perez, A. Piktus et al., "Retrievalaugmented generation for knowledge-intensive NLP tasks," in NeurIPS, vol. 33, pp. 9459-9474, 2020.
  12. Q. Wu, G. Bansal, J. Zhang et al., "AutoGen: Enabling next-gen LLM applications via multia gentconversat ion, "arXivpreprintarXiv:2308.08155, 2023.
  13. OpenAI, "GPT-4 technical report," arXiv preprint arXiv:2303.08774, 2023.
  14. Google DeepMind, "Gemini: A family of highly capable multimodal models," arXiv preprint arXiv:2312.11805, 2023.
  15. Meta AI, "Llama 3: Open foundation and fine-tuned chat models," Meta AI Blog, 2024.
  16. Lu, C. Lu, R. T. Q. Chen et al., "The AI Scientist: Towards fully automated open-ended scientific discovery," arXiv preprint arXiv:2408.06292, 2024.
  17. S. Jimenez, J. Yang, H. Wettig et al., "SWE-bench: Can language models resolve real-world GitHub issues?," in ICLR. 2024.
  18. T. Liu, X. Zhang, B. Guo et al., "AgentBench: Evaluating LLMs as agents," in ICLR, 2024.
  19. E. Hu, Y. Shen, P. Wallis et al., "LoRA: Low-rank adaptation of large language models," in ICLR, 2022.
  20. M. Yang et al., "HotpotQA: A dataset for diverse, explainable multi-hop question answering," EMNLP, 2018.
  21. Hendrycks et al., "Measuring mathematical problem solving with the MATH dataset," NeurIPS, 2021.
  22. M. Joshi et al., "TriviaQA: A reading comprehension dataset over trivia questions," ACL, 2017.
Scroll to Top