Diminishing Returns: LLM Accuracy vs. Computational Cost on Constrained Hardware

• Gustavo Fernandes Luz UFAPE
• Tiago Buarque Assunção de Carvalho UFAPE

DOI: https://doi.org/10.5753/eniac.2026.29707

Resumo

Large Language Models (LLMs) have become increasingly vital across diverse fields. This study investigates the trade-off between accuracy and computational cost in language models when operating under hardware constraints. While mainstream research relies heavily on parametric scaling, our evaluation of thirteen models on the MMLU benchmark reveals diminishing marginal returns. Specifically, increasing parameter counts yields progressively smaller accuracy gains while drastically increasing execution time. We conclude that continuous structural scaling is inefficient for resource-constrained environments.

Referências

Agrawal, R., Asrani, M., Youssef, H., and Narayan, A. (2025). Scmrag: Self-corrective multihop retrieval augmented generation system for llm agents. In Proceedings of the 24th International Conference on Autonomous Agents and Multiagent Systems (AAMAS), Detroit, MI, USA. International Foundation for Autonomous Agents and Multiagent Systems.

Agustiani, A. D., Putri, S. M., Hutahaean, J., Sholahuddin, M. R., Alifi, M. R., and Hodijah, A. (2026). Evaluating rag performance on small language models for low-resource devices through chunking and retrieval methods. Jurnal Online Informatika, 11(1):45–56.

Chowdhury, M. D. T. B. Z., Hossain, M., and Islam, M. R. (2025). Durghotona gpt: A web scraping and large language model based framework to generate road accident dataset automatically in bangladesh. arXiv preprint, abs/2504.21025.

Desislavov, R., Martínez-Plumed, F., and Hernández-Orallo, J. (2023). Trends in ai inference energy consumption: Beyond the performance-vs-parameter laws of deep learning. Sustainable Computing: Informatics and Systems, 38:100857.

Filho, J. F. and Fernandes, S. (2025). Do slm agents screen papers as well as llms? In Anais do XXII Encontro Nacional de Inteligência Artificial e Computacional, pages 592–603, Porto Alegre, RS, Brasil. SBC.

Ganesh, P., Goel, K., and McCallum, A. (2023). Small language models are also few-shot learners. arXiv preprint arXiv:2306.07930.

Gupta, D., Suganya, R., Bhoji, S., Geetha, S., Arora, M., and Manjula, D. (2024). Enhanced research on legal data extraction and document analysis tool. In Proceedings of the 8th International Conference on I-SMAC (IoT in Social, Mobile, Analytics and Cloud), pages 756–763. IEEE.

Hachicha, M., Omezine, M. A., Zghal, M. K., Riahi, M. H., and Ncib, L. (2024). Recommendation module based on web scraping and llm models with the rag technique. In Proceedings of the Artificial Intelligence Revolutions (AIR). IEEE.

Hoffmann, J. et al. (2022). Training compute-optimal large language models. arXiv preprint arXiv:2203.15556.

Huang, Z., Tang, J., Karir, M., Liu, M., and Sarabi, A. (2024). Analyzing corporate privacy policies using ai chatbots. In Proceedings of the 2024 ACM on Internet Measurement Conference (IMC ’24), pages 505–515, Madrid, Spain. Association for Computing Machinery.

Jamil, H., Noorosmawie, A. B., Rabu, H. W., and Razak, L. A. (2025). From archives to ai: Residential property data across three decades in brunei darussalam. Data in Brief, 60:111505.

Jervas, S., Jacob, C., Sundar, S., and Varma, D. P. (2025). Context-driven chatbot development: Leveraging zephyr-7b with rag for improved response accuracy. In Proceedings of the Emerging Technologies for Intelligent Systems Conference. Conference paper, Feb. 2025.

Kanataria, N., Goel, P., Patel, K. P., Patel, K., Patel, H. N., and Garg, D. (2024). Ragenhanced large language model for intelligent assistance from web-scraped data. In Proceedings of the 9th International Conference on Communication and Electronics Systems (ICCES). IEEE.

Kaplan, J., McCandlish, S., Henighan, T., et al. (2020). Scaling laws for neural language models. arXiv preprint arXiv:2001.08361.

Kumar, K. S., Srihari, P., and Raman, C. J. (2024). Ai for career growth: Advanced resume analysis and linkedin scraping for personalized job recommendations. In Proceedings of the 2nd International Conference on Self Sustainable Artificial Intelligence Systems (ICSSAS-2024), pages 1287–1293, Chennai, India. IEEE.

Kumbar, P., Sharma, S., Batra, I., Malik, A., and Ashfaq, F. (2024). Storyspinner: Ai based technique to narrate the web links. In Proceedings of the International Conference on Emerging Trends in Networks and Computer Communications (ETNCC). IEEE.

Liu, H., Meng, Y., Zhong, Y., Zhao, C., and Han, J. (2024). Small language models: Survey, measurements, and insights. arXiv preprint arXiv:2401.03720.

Oliveira, J., Rufino, N., Levcovitz, S., Utino, M., Lobato, F., Turine, M., Rezende, S., Junior, A. J., and Marcacini, R. (2025). Agents4gov: Privacy-preserving web agents using open llms for public sector tasks. In Anais do XXII Encontro Nacional de Inteligência Artificial e Computacional, pages 580–591, Porto Alegre, RS, Brasil. SBC.

Porlou, C., Makrynioti, M., Alexiadis, A., Stavropoulos, G., Pantelis, G., Votis, K., and Tzovaras, D. (2024). Optimizing an llm prompt for accurate data extraction from firearm-related listings in dark web marketplaces. In Proceedings of the IEEE International Conference on Big Data. IEEE.

Prucs, Á., Csutora, M., Antal, M., and Marosi, M. (2025). Compute-accuracy pareto frontiers for open-source reasoning large language models. arXiv preprint.

Pushpalatha, M. and Aravindan, M. S. (2025). Comparative analysis of web scraping methodologies using generative ai. In Proceedings of the 6th IEEE International Conference on Recent Advances in Information Technology (RAIT). IEEE.

Rathod, M., Yadav, A., Phutane, S., and Bhise, S. (2024). Comment compass: An approach to analyze youtube comments through web scraping, sentiment analysis, and generative ai for actionable insights. In Proceedings of the 2024 First International Conference on Pioneering Developments in Computer Science & Digital Technologies (IC2SDT 2024). IEEE.

Saad, M. and Qawaqneh, Z. (2024). Closed domain question-answering techniques in an institutional chatbot. In Proceedings of the 2024 4th International Conference on Electrical, Computer, Communications and Mechatronics Engineering (ICECCME 2024), Male, Maldives.

Salim, M. S., Hossain, S. I., Jalal, T., Bose, D. K., and Basher, M. J. I. (2025). Llm based qa chatbot builder: A generative ai-based chatbot builder for question answering. SoftwareX, 29:102029.

Sampaio, T., Oliveira, P. F., and Matos, P. J. (2024). Development of a chatbot to support an university institutional website. In Proceedings of the 2024 International Conference on Engineering and Emerging Technologies (ICEET).

Senapati, A., Ananya, J., Jha, A., Mahishi, A., and Srinivas, K. S. (2024). Driving innovation: Creating a dataset for automotive ad campaign analysis. In Proceedings of the 2024 First International Conference on Pioneering Developments in Computer Science & Digital Technologies (IC2SDT), pages 273–278. IEEE.

Sriram, G., Suji, A. R., Nivitha, K., Rahul, S., Nallusamy, P., and Arun, A. (2024). Scrapetalk: Chatbot conversation with web scraped insights. In Proceedings of the International Conference on Integration of Emerging Technologies for the Digital World (ICIETDW). IEEE.

Sufi, F. and Alsulami, M. (2025). Disaster in the headlines: Quantifying narrative variation in global news using topic modeling and statistical inference. Mathematics, 13(13):2049.

Tan, D., Huang, Y., Liu, M., Li, Z., Wu, X., and Huang, C. (2025). Identification of online health information using large pretrained language models: Mixed methods study. Journal of Medical Internet Research, 27:e70733.

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., and Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30.

Welch, B. L. (1947). The generalization of ‘student’s’ problem when several different population varlances are involved. Biometrika, 34(1-2):28–35.

Zhao, Q., Huang, Y., Lv, T., Cui, L., Wei, F., et al. (2024). Mmlu-cf: A contamination-free multi-task language understanding benchmark. arXiv preprint arXiv:2412.15194.