Ajuste Automático de Parâmetros Bayesiano com Persistência de Contexto para Operadores Semânticos
Resumo
O Multi-Armed Bandit (MAB), guiado por intervalos de confiança, tem sido empregado na otimização de planos de execução em sistemas baseados em operadores semânticos sobre LLMs. Entretanto, essas abordagens normalmente reinicializam suas estimativas a cada nova execução, sem reutilizar o conhecimento acumulado em execuções anteriores. Este trabalho reformula a amostragem como um procedimento Bayesiano conduzido por Thompson Sampling, propondo o algoritmo CXTBAY, que persiste distribuições a posteriori entre execuções e interpola priors informativas, para operadores nunca amostrados, a partir de um Processo Gaussiano sobre embeddings de operadores. Comparado ao MAB, o CXTBAY conseguiu encontrar o mesmo plano final 85% das vezes, encontrando planos de qualidade superior nos demais casos. Isto foi feito com até 2,3× menos tempo e recursos gastos no processo de otimização.Referências
Agrawal, R. (1995). The Continuum-Armed Bandit Problem. SIAM J. Control Optim., 33(6):1926–1951.
Anderson, E., Fritz, J., Lee, A., Li, B., Lindblad, M., Lindeman, H., Meyer, A., Parmar, P., Ranade, T., Shah, M. A., Sowell, B., Tecuci, D., Thapliyal, V., and Welsh, M. (2025). The design of an LLM-powered unstructured analytics system. In Proceedings of the Conference on Innovative Data Systems Research, CIDR.
Berk, R. H. (1996). Continuous univariate distributions, volume 2.
Delecate, K., Van der Straeten, S., Luyckx, K., and Demeester, T. (2023). BioDEX: Large-scale biomedical adverse drug event extraction for real-world pharmacovigilance. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 13429–13446, Singapore. Association for Computational Linguistics.
Hendrycks, D., Burns, C., Chen, A., and Ball, S. (2021). CUAD: An Expert-Annotated NLP Dataset for Legal Contract Review. In Advances in Neural Information Processing Systems (NeurIPS), volume 34, pages 28479–28492.
Kwon, W. et al. (2023). Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGCOMM Conference.
Liu, C. et al. (2025). Palimpzest: Declarative data processing for ai pipelines. In Proceedings of the Conference on Innovative Data Systems Research (CIDR).
Patel, P. et al. (2025). Lotus: Enabling semantic queries with large language models over structured and unstructured data. Proceedings of the VLDB Endowment (PVLDB).
Russo, D., Roy, B. V., Kazerouni, A., Osband, I., and Wen, Z. (2018). A tutorial on thompson sampling. Foundations and Trends in Machine Learning, 11(1):1–96.
Russo, M. et al. (2026). Abacus: A cost-based optimizer for semantic operator systems. Proceedings of the VLDB Endowment (PVLDB).
Sanmartino, G., Urban, M., Papotti, P., and Binnig, C. (2026). The stretto execution engine for llm-augmented data systems. arXiv preprint arXiv:2602.04430.
Shahriari, B., Swersky, K., Wang, Z., Adams, R. P., and de Freitas, N. (2016). Taking the human out of the loop: A review of bayesian optimization. Proceedings of the IEEE, 104(1):148–175.
Shankar, S., Chambers, T., Shah, T., Parameswaran, A. G., and Wu, E. (2025). DocETL: Agentic query rewriting and evaluation for complex document processing. Proceedings of the VLDB Endowment, 18(9):3035–3048.
Thakur, N., Reimers, N., Ruckl’e, A., Srivastava, A., and Gurevych, I. (2021). BEIR: A Heterogenous Benchmark for Zero-shot Evaluation of Information Retrieval Models. ArXiv, abs/2104.08663.
Wang, Y., Audibert, J.-y., and Munos, R. (2008). Algorithms for Infinitely Many-Armed Bandits. In Koller, D., Schuurmans, D., Bengio, Y., and Bottou, L., editors, Advances in Neural Information Processing Systems, volume 21. Curran Associates, Inc.
William, R. T. et al. (1933). On the likelihood that one unknown probability exceeds another in view of the evidence of two samples. Biometrika, 25(3/4):285–294.
Zhong, Z., Chueng, W. C., and Tan, V. Y. (2021). Thompson sampling algorithms for cascading bandits. Journal of Machine Learning Research, 22(218):1–66.
Anderson, E., Fritz, J., Lee, A., Li, B., Lindblad, M., Lindeman, H., Meyer, A., Parmar, P., Ranade, T., Shah, M. A., Sowell, B., Tecuci, D., Thapliyal, V., and Welsh, M. (2025). The design of an LLM-powered unstructured analytics system. In Proceedings of the Conference on Innovative Data Systems Research, CIDR.
Berk, R. H. (1996). Continuous univariate distributions, volume 2.
Delecate, K., Van der Straeten, S., Luyckx, K., and Demeester, T. (2023). BioDEX: Large-scale biomedical adverse drug event extraction for real-world pharmacovigilance. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 13429–13446, Singapore. Association for Computational Linguistics.
Hendrycks, D., Burns, C., Chen, A., and Ball, S. (2021). CUAD: An Expert-Annotated NLP Dataset for Legal Contract Review. In Advances in Neural Information Processing Systems (NeurIPS), volume 34, pages 28479–28492.
Kwon, W. et al. (2023). Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGCOMM Conference.
Liu, C. et al. (2025). Palimpzest: Declarative data processing for ai pipelines. In Proceedings of the Conference on Innovative Data Systems Research (CIDR).
Patel, P. et al. (2025). Lotus: Enabling semantic queries with large language models over structured and unstructured data. Proceedings of the VLDB Endowment (PVLDB).
Russo, D., Roy, B. V., Kazerouni, A., Osband, I., and Wen, Z. (2018). A tutorial on thompson sampling. Foundations and Trends in Machine Learning, 11(1):1–96.
Russo, M. et al. (2026). Abacus: A cost-based optimizer for semantic operator systems. Proceedings of the VLDB Endowment (PVLDB).
Sanmartino, G., Urban, M., Papotti, P., and Binnig, C. (2026). The stretto execution engine for llm-augmented data systems. arXiv preprint arXiv:2602.04430.
Shahriari, B., Swersky, K., Wang, Z., Adams, R. P., and de Freitas, N. (2016). Taking the human out of the loop: A review of bayesian optimization. Proceedings of the IEEE, 104(1):148–175.
Shankar, S., Chambers, T., Shah, T., Parameswaran, A. G., and Wu, E. (2025). DocETL: Agentic query rewriting and evaluation for complex document processing. Proceedings of the VLDB Endowment, 18(9):3035–3048.
Thakur, N., Reimers, N., Ruckl’e, A., Srivastava, A., and Gurevych, I. (2021). BEIR: A Heterogenous Benchmark for Zero-shot Evaluation of Information Retrieval Models. ArXiv, abs/2104.08663.
Wang, Y., Audibert, J.-y., and Munos, R. (2008). Algorithms for Infinitely Many-Armed Bandits. In Koller, D., Schuurmans, D., Bengio, Y., and Bottou, L., editors, Advances in Neural Information Processing Systems, volume 21. Curran Associates, Inc.
William, R. T. et al. (1933). On the likelihood that one unknown probability exceeds another in view of the evidence of two samples. Biometrika, 25(3/4):285–294.
Zhong, Z., Chueng, W. C., and Tan, V. Y. (2021). Thompson sampling algorithms for cascading bandits. Journal of Machine Learning Research, 22(218):1–66.
Publicado
03/11/2026
Como Citar
VINÍCIUS, Arthur; REGI, Luiza; ALEXANDRE, Rafael; BARREIROS JR., Willian; TEODORO, George.
Ajuste Automático de Parâmetros Bayesiano com Persistência de Contexto para Operadores Semânticos. In: SIMPÓSIO EM SISTEMAS COMPUTACIONAIS DE ALTO DESEMPENHO (SSCAD), 27. , 2026, Natal/RN.
Anais [...].
Porto Alegre: Sociedade Brasileira de Computação,
2026
.
p. 497-509.
DOI: https://doi.org/10.5753/sscad.2026.34966.
