Mapping Researchers to Strategic Themes with Language Models via Reasoning- and Reward-Guided Fine-Tuning
Resumo
Mapping researcher profiles to strategic topics is a key challenge in the Brazilian National Graduate System (SNPG), as it supports public policy planning and research funding allocation. To address this problem, this paper proposes TORNADO, a TwO-stage ReasoNing-rewArD Optimization method that combines reasoning supervision and reward optimization. Its main contribution is a distance-aware reward that models the ordinal structure of affinity labels by penalizing severe errors more than adjacent-class mistakes. Experiments on the LeanDL-HPC dataset show that TORNADO outperforms zero-shot, few-shot, and conventional supervised fine-tuning baselines, providing an efficient and scalable alternative for strategic decision-making within the SNPG.Referências
Al Faraby, S., Romadhony, A., and Adiwijaya (2024). Analysis of llms for educational question classification and generation. Computers and Education: Artificial Intelligence.
CAPES (2023). Plano nacional de pós-graduação – PNPG 2024–2028: versão preliminar para consulta pública. Technical report, Ministério da Educação / CAPES, Brasília.
Chang, Y., Wang, X., Wang, J., Wu, Y., Yang, L., Zhu, K., Chen, H., Yi, X., Wang, C., Wang, Y., et al. (2024). A survey on evaluation of large language models. ACM transactions on intelligent systems and technology, 15(3):1–45.
Gôlo, M. P. S. and Utino, M. Y. R. (2025). One-class lightweight interpretable filtering for academic profiles and strategic themes affinity. In International Symposium on Computer Architecture and High Performance Computing Workshops, pages 147–154.
Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., et al. (2025). Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.
Liu, X., Zheng, Y., Du, Z., Ding, M., Qian, Y., Yang, Z., and Tang, J. (2023). Gpt understands, too. AI Open.
Matsubura, E. T., Borin, E., Cirne, W., Cáceres, E. N., Marcacini, R., Nogueira, B., Silva, D. F., de Andrade Silva, J., Turine, M. A. S., de Souza, M. C., et al., editors (2025). Workshop on Lightweight and Efficient Deep Learning in HPC Environments.
Miranda, A. C. D., Utino, M. Y. R., Gôlo, M. P. S., Dos Santos, M. A. A., and de Souza, M. C. (2025). Reducing costs in large-scale classification: A hybrid bert–llm strategy. In Int. Symp. on Computer Architecture and High Performance Computing Workshops.
Mustafa, A., Naseem, U., and Rahimi Azghadi, M. (2026). Can reasoning llms enhance clinical document classification? Health and Technology, 16.
Sakai, H. and Lam, S. S. (2026). Large language models for health care text classification: Systematic review. JMIR AI, 5:e79202.
Sakiyama, K. M., Fujimoto, M. L., Santin, R. V., and Rezende, S. O. (2025). Leandl hpc challenge 2025: Applying large-scale model adaptation techniques. In Int. Symp. on Computer Architecture and High Performance Computing Workshops, pages 139–146.
Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y., et al. (2024). Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300.
Turine, M. A. S., Souto Maior, A. P., and Marcacini, R. M. (2025). National agenda for the training of higher-level personnel. Technical Report 1, CAPES – Coordination for the Improvement of Higher Education Personnel, Ministry of Education, Brazil.
United Nations (2015). Transforming our world: The 2030 agenda for sustainable development. Technical report, United Nations. Accessed: Apr. 20, 2026.
Utino, M. Y. R. and Gôlo, M. P. S. (2025). Efficient filtering with bert embeddings for researcher–topic affinity prediction in hpc pipelines. In International Symposium on Computer Architecture and High Performance Computing Workshops, pages 155–162.
Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q. V., Zhou, D., et al. (2022). Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837.
CAPES (2023). Plano nacional de pós-graduação – PNPG 2024–2028: versão preliminar para consulta pública. Technical report, Ministério da Educação / CAPES, Brasília.
Chang, Y., Wang, X., Wang, J., Wu, Y., Yang, L., Zhu, K., Chen, H., Yi, X., Wang, C., Wang, Y., et al. (2024). A survey on evaluation of large language models. ACM transactions on intelligent systems and technology, 15(3):1–45.
Gôlo, M. P. S. and Utino, M. Y. R. (2025). One-class lightweight interpretable filtering for academic profiles and strategic themes affinity. In International Symposium on Computer Architecture and High Performance Computing Workshops, pages 147–154.
Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., et al. (2025). Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948.
Liu, X., Zheng, Y., Du, Z., Ding, M., Qian, Y., Yang, Z., and Tang, J. (2023). Gpt understands, too. AI Open.
Matsubura, E. T., Borin, E., Cirne, W., Cáceres, E. N., Marcacini, R., Nogueira, B., Silva, D. F., de Andrade Silva, J., Turine, M. A. S., de Souza, M. C., et al., editors (2025). Workshop on Lightweight and Efficient Deep Learning in HPC Environments.
Miranda, A. C. D., Utino, M. Y. R., Gôlo, M. P. S., Dos Santos, M. A. A., and de Souza, M. C. (2025). Reducing costs in large-scale classification: A hybrid bert–llm strategy. In Int. Symp. on Computer Architecture and High Performance Computing Workshops.
Mustafa, A., Naseem, U., and Rahimi Azghadi, M. (2026). Can reasoning llms enhance clinical document classification? Health and Technology, 16.
Sakai, H. and Lam, S. S. (2026). Large language models for health care text classification: Systematic review. JMIR AI, 5:e79202.
Sakiyama, K. M., Fujimoto, M. L., Santin, R. V., and Rezende, S. O. (2025). Leandl hpc challenge 2025: Applying large-scale model adaptation techniques. In Int. Symp. on Computer Architecture and High Performance Computing Workshops, pages 139–146.
Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y., et al. (2024). Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300.
Turine, M. A. S., Souto Maior, A. P., and Marcacini, R. M. (2025). National agenda for the training of higher-level personnel. Technical Report 1, CAPES – Coordination for the Improvement of Higher Education Personnel, Ministry of Education, Brazil.
United Nations (2015). Transforming our world: The 2030 agenda for sustainable development. Technical report, United Nations. Accessed: Apr. 20, 2026.
Utino, M. Y. R. and Gôlo, M. P. S. (2025). Efficient filtering with bert embeddings for researcher–topic affinity prediction in hpc pipelines. In International Symposium on Computer Architecture and High Performance Computing Workshops, pages 155–162.
Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q. V., Zhou, D., et al. (2022). Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837.
Publicado
19/10/2026
Como Citar
MIRANDA, Augusto C. D.; GÔLO, Marcos P. S.; SANTOS, Marcela A. A. dos; MARCACINI, Ricardo M.; TURINE, Marcelo A. S.; SOUZA, Mariana C. de.
Mapping Researchers to Strategic Themes with Language Models via Reasoning- and Reward-Guided Fine-Tuning. In: ENCONTRO NACIONAL DE INTELIGÊNCIA ARTIFICIAL E COMPUTACIONAL (ENIAC), 23. , 2026, Cuiabá/MT.
Anais [...].
Porto Alegre: Sociedade Brasileira de Computação,
2026
.
p. 1136-1147.
ISSN 2763-9061.
DOI: https://doi.org/10.5753/eniac.2026.30884.
