Architecture over Scale: Multi-Agent Reasoning for ENEM Mathematics
Resumo
Este artigo avalia uma arquitetura multiagente leve para raciocínio matemático em modelos de linguagem pequenos. Em 301 questões textuais e autocontidas de matemática do ENEM, comparamos o raciocínio direto, duas configurações multiagente de 20B e a inferência direta em um modelo de 120B da mesma família. O sistema multiagente verificado de 20B alcança 97,7% de acurácia, igualando o modelo de 120B e superando os 91,7% do modelo de 20B direto. Embora aumente o uso de tokens e o custo, a arquitetura permite obter desempenho de modelos maiores com um modelo adequado à execução local.Referências
Almeida, T. S., Laitz, T., Bonás, G. K., and Nogueira, R. (2023). BLUEX: A benchmark based on Brazilian leading universities entrance exams. arXiv preprint arXiv:2307.05410.
Belcak, P., Heinrich, G., Diao, S., Fu, Y., Dong, X., Muralidharan, S., Lin, Y. C., and Molchanov, P. (2025). Small language models are the future of agentic AI. arXiv preprint arXiv:2506.02153.
Besta, M., Blach, N., Kubicek, A., Gerstenberger, R., Podstawski, M., Gianinazzi, L., Gajda, J., Lehmann, T., Niewiadomski, H., Nyczyk, P., and Hoefler, T. (2024). Graph of thoughts: Solving elaborate problems with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 17682–17690.
Chen, W., Ma, X., Wang, X., and Cohen, W. W. (2023). Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks. Transactions on Machine Learning Research.
Cobbe, K., Kosaraju, V., Bavarian, M., Chen, M., Jun, H., Kaiser, L., Plappert, M., Tworek, J., Hilton, J., Nakano, R., Hesse, C., and Schulman, J. (2021). Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.
Gao, L., Madaan, A., Zhou, S., Alon, U., Liu, P., Yang, Y., Callan, J., and Neubig, G. (2022). Pal: Program-aided language models. arXiv preprint arXiv:2211.10435.
Gou, Z., Shao, Z., Gong, Y., Shen, Y., Yang, Y., Huang, M., Duan, N., and Chen, W. (2024). ToRA: A tool-integrated reasoning agent for mathematical problem solving. In International Conference on Learning Representations.
Groq, Inc. (2026a). GroqCloud: Inference-as-a-service platform. [link]. Acessado em 2026.
Groq, Inc. (2026b). GroqCloud pricing. [link]. Acessado em 2026.
Hendrycks, D., Burns, C., Kadavath, S., Arora, A., Basart, S., Tang, E., Song, D., and Steinhardt, J. (2021). Measuring mathematical problem solving with the MATH dataset. In Advances in Neural Information Processing Systems, volume 34.
Huang, J., Chen, X., Mishra, S., Zheng, H. S., Yu, A. W., Song, X., and Zhou, D. (2024). Large language models cannot self-correct reasoning yet. In International Conference on Learning Representations.
Nunes, D., Primi, R., Pires, R., Lotufo, R., and Nogueira, R. (2023). Evaluating GPT-3.5 and GPT-4 models on Brazilian university admission exams. arXiv preprint arXiv:2303.17003.
OpenAI (2025). GPT-OSS-20B and GPT-OSS-120B: Open-weight reasoning models. arXiv preprint arXiv:2508.12461.
Pires, R., de Souza, T. S., Ribeiro, P. H., and Lotufo, R. (2023). Evaluating GPT-4’s vision capabilities on Brazilian university admission exams. arXiv preprint arXiv:2311.14169.
Schick, T., Dwivedi-Yu, J., Dessi, R., Raileanu, R., Lomeli, M., Hambro, E., Zettlemoyer, L., Cancedda, N., and Scialom, T. (2023). Toolformer: Language models can teach themselves to use tools. In Advances in Neural Information Processing Systems, volume 36, pages 68539–68551.
Shinn, N., Cassano, F., Gopinath, A., Narasimhan, K., and Yao, S. (2023). Reflexion: Language agents with verbal reinforcement learning. In Advances in Neural Information Processing Systems, volume 36, pages 8634–8652.
Toh, V. et al. (2024). Not all votes count! programs as verifiers improve self-consistency of language models for math reasoning. arXiv preprint arXiv:2410.12608.
Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E. H., Le, Q. V., and Zhou, D. (2022). Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems (NeurIPS).
Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., and Cao, Y. (2023). ReAct: Synergizing reasoning and acting in language models. In International Conference on Learning Representations.
Zhou, A., Wang, K., Lu, Z., Shi, W., Luo, S., Qin, Z., Lu, S., Jia, A., Song, L., Zhan, M., and Li, H. (2023). Solving challenging math word problems using GPT-4 code interpreter with code-based self-verification. arXiv preprint arXiv:2308.07921.
Żywot, A., Chen, X., and de Rijke, M. (2026). Can small agent collaboration beat a single big LLM? arXiv preprint arXiv:2601.11327.
Belcak, P., Heinrich, G., Diao, S., Fu, Y., Dong, X., Muralidharan, S., Lin, Y. C., and Molchanov, P. (2025). Small language models are the future of agentic AI. arXiv preprint arXiv:2506.02153.
Besta, M., Blach, N., Kubicek, A., Gerstenberger, R., Podstawski, M., Gianinazzi, L., Gajda, J., Lehmann, T., Niewiadomski, H., Nyczyk, P., and Hoefler, T. (2024). Graph of thoughts: Solving elaborate problems with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 17682–17690.
Chen, W., Ma, X., Wang, X., and Cohen, W. W. (2023). Program of thoughts prompting: Disentangling computation from reasoning for numerical reasoning tasks. Transactions on Machine Learning Research.
Cobbe, K., Kosaraju, V., Bavarian, M., Chen, M., Jun, H., Kaiser, L., Plappert, M., Tworek, J., Hilton, J., Nakano, R., Hesse, C., and Schulman, J. (2021). Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168.
Gao, L., Madaan, A., Zhou, S., Alon, U., Liu, P., Yang, Y., Callan, J., and Neubig, G. (2022). Pal: Program-aided language models. arXiv preprint arXiv:2211.10435.
Gou, Z., Shao, Z., Gong, Y., Shen, Y., Yang, Y., Huang, M., Duan, N., and Chen, W. (2024). ToRA: A tool-integrated reasoning agent for mathematical problem solving. In International Conference on Learning Representations.
Groq, Inc. (2026a). GroqCloud: Inference-as-a-service platform. [link]. Acessado em 2026.
Groq, Inc. (2026b). GroqCloud pricing. [link]. Acessado em 2026.
Hendrycks, D., Burns, C., Kadavath, S., Arora, A., Basart, S., Tang, E., Song, D., and Steinhardt, J. (2021). Measuring mathematical problem solving with the MATH dataset. In Advances in Neural Information Processing Systems, volume 34.
Huang, J., Chen, X., Mishra, S., Zheng, H. S., Yu, A. W., Song, X., and Zhou, D. (2024). Large language models cannot self-correct reasoning yet. In International Conference on Learning Representations.
Nunes, D., Primi, R., Pires, R., Lotufo, R., and Nogueira, R. (2023). Evaluating GPT-3.5 and GPT-4 models on Brazilian university admission exams. arXiv preprint arXiv:2303.17003.
OpenAI (2025). GPT-OSS-20B and GPT-OSS-120B: Open-weight reasoning models. arXiv preprint arXiv:2508.12461.
Pires, R., de Souza, T. S., Ribeiro, P. H., and Lotufo, R. (2023). Evaluating GPT-4’s vision capabilities on Brazilian university admission exams. arXiv preprint arXiv:2311.14169.
Schick, T., Dwivedi-Yu, J., Dessi, R., Raileanu, R., Lomeli, M., Hambro, E., Zettlemoyer, L., Cancedda, N., and Scialom, T. (2023). Toolformer: Language models can teach themselves to use tools. In Advances in Neural Information Processing Systems, volume 36, pages 68539–68551.
Shinn, N., Cassano, F., Gopinath, A., Narasimhan, K., and Yao, S. (2023). Reflexion: Language agents with verbal reinforcement learning. In Advances in Neural Information Processing Systems, volume 36, pages 8634–8652.
Toh, V. et al. (2024). Not all votes count! programs as verifiers improve self-consistency of language models for math reasoning. arXiv preprint arXiv:2410.12608.
Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E. H., Le, Q. V., and Zhou, D. (2022). Chain-of-thought prompting elicits reasoning in large language models. In Advances in Neural Information Processing Systems (NeurIPS).
Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., and Cao, Y. (2023). ReAct: Synergizing reasoning and acting in language models. In International Conference on Learning Representations.
Zhou, A., Wang, K., Lu, Z., Shi, W., Luo, S., Qin, Z., Lu, S., Jia, A., Song, L., Zhan, M., and Li, H. (2023). Solving challenging math word problems using GPT-4 code interpreter with code-based self-verification. arXiv preprint arXiv:2308.07921.
Żywot, A., Chen, X., and de Rijke, M. (2026). Can small agent collaboration beat a single big LLM? arXiv preprint arXiv:2601.11327.
Publicado
19/10/2026
Como Citar
FERREIRA, Paulo Victor C.; SILVEIRA, Erica C.; SANTOS, Guilherme L. dos; B. NETO, Francisco Leudes; OLIVEIRA, Marcos A. de.
Architecture over Scale: Multi-Agent Reasoning for ENEM Mathematics. In: ENCONTRO NACIONAL DE INTELIGÊNCIA ARTIFICIAL E COMPUTACIONAL (ENIAC), 23. , 2026, Cuiabá/MT.
Anais [...].
Porto Alegre: Sociedade Brasileira de Computação,
2026
.
p. 3310-3321.
ISSN 2763-9061.
DOI: https://doi.org/10.5753/eniac.2026.31329.
