SLLMix: Uma Ferramenta Interativa para Explorar o Trade-off entre SLMs e LLMs

• Claudio M. V. de Andrade UFMG
• Elisa T. Albergaria UFSJ
• Leonardo C. D. Rocha UFSJ
• Marcos A. Gonçalves UFMG

DOI: https://doi.org/10.5753/webmedia_estendido.2026.34624

Resumo

Automatic Text Classification (ATC) is a central task in Natural Language Processing (NLP), commonly addressed with Transformer-based Small Language Models (SLMs) that offer competitive effectiveness at relatively low computational cost. Large Language Models (LLMs), in turn, provide additional language understanding and prompting capabilities, but at substantially higher inference costs. Selectively combining these model families raises a practical question: when is it worth routing instances from an SLM to an LLM? In this work, we present SLLMix (Small Language Models + Large Language Models Mix), an interactive web-based tool for configuring, executing, and analyzing selective SLM–LLM combinations for text classification. SLLMix enables users to upload datasets, choose SLM and LLM models, define prompts, and explore how confidence thresholds affect both the number of instances routed to the LLM and the resulting effectiveness–efficiency trade-off. The tool integrates classification metrics, efficiency measurements, interactive visualizations, and qualitative inspection of cases in which the LLM corrects SLM errors or both models fail. By combining quantitative and qualitative analysis in a single environment, SLLMix supports reproducible and extensible investigation of when SLM–LLM combinations are beneficial.

Palavras-chave: Automatic Text Classification, Small Language Models, Large Language Models, Model Combination, Interactive Evaluation, Web-based Tool

Referências

Claudio Moisés Valiense De Andrade, Washington Cunha, Guilherme Fonseca, Ana Clara Souza Pagano, Luana De Castro Santos, Adriana Silvina Pagano, Leonardo Chaves Dutra Da Rocha, and Marcos André Gonçalves. 2024. Explaining the Hardest Errors of Contextual Embedding Based Classifiers. In Proceedings of the 28th Conference on Computational Natural Language Learning, Libby Barak and Malihe Alikhani (Eds.). Association for Computational Linguistics, Miami, FL, USA, 419–434.

Claudio Moisés Valiense de Andrade, Washington Cunha, Davi Reis, Celso França, Wasterman Ávila Apolinário, Luana de Castro Santos, Adriana Silvina Pagano, Leonardo Chaves Dutra da Rocha, and Marcos André Gonçalves. 2026. Call My Big Sibling (CMBS) – A Confidence-Based Strategy Leveraging Instance Selection to Combine Small and Large Language Models for Cost-Effective Text Classification. Journal of the Brazilian Computer Society 32, 1 (May 2026), 1233–1249.

Lingjiao Chen, Matei Zaharia, and James Zou. 2023. FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance. arXiv:2305.05176 [cs.LG]

Evandro Cunha, Gabriel Magno, Marcos André Gonçalves, César Cambraia, and Virgilio Almeida. 2014. He Votes or She Votes? Female and Male Discursive Strategies in Twitter Political Hashtags. PLOS ONE 9, 1 (2014), e87041. DOI: 10.1371/journal.pone.0087041

Claudio M.V. de Andrade, Fabiano M. Belém, Washington Cunha, Celso França, Felipe Viegas, Leonardo Rocha, and Marcos André Gonçalves. 2023. On the class separability of contextual embeddings representations – or “The classifier does not matter when the (text) representation is so good!”. Information Processing Management 60, 4 (2023), 103336.

Percy Liang et al. 2023. Holistic Evaluation of Language Models. arXiv:2211.09110 [cs.CL] [link]

Vijay Arya et al. 2019. One Explanation Does Not Fit All: A Toolkit and Taxonomy of AI Explainability Techniques.

David Fernandes, Edleno S. de Moura, Berthier Ribeiro-Neto, Altigran S. da Silva, and Marcos André Gonçalves. 2007. Computing Block Importance for Searching on Web Sites. In Proceedings of the 16th ACM Conference on Information and Knowledge Management (CIKM ’07). ACM, 165–174.

Google Cloud. 2025. Cloud Natural Language. [link]. Accessed: 2026-07-07.

Guillem Ramírez, Alexandra Birch, and Ivan Titov. 2024. Optimising Calls to Large Language Models with Uncertainty-Based Two-Tier Selection. arXiv:2405.02134 [cs.CL]

Felipe Viegas, Washington Cunha, Christian Gomes, Antônio Pereira, Leonardo Rocha, and Marcos André Gonçalves. 2020. CluHTM – Semantic Hierarchical Topic Modeling based on CluWords. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. Association for Computational Linguistics, 8138–8150.

James Wexler, Mahima Pushkarna, Tolga Bolukbasi, Martin Wattenberg, Fernanda B. Viégas, and Jimbo Wilson. 2019. The What-If Tool: Interactive Probing of Machine Learning Models. CoRR abs/1907.04135 (2019). arXiv:1907.04135

Thomas Wolf, Lysandre Debut, Victor Sanh, et al. 2020. Transformers: State-of-the-Art Natural Language Processing. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations. 38–45.