CulturaQuant: Quantization and Brazilian Cultural Knowledge in Small Language Models

• Cristhian Kapelinski UNIPAMPA https://orcid.org/0009-0005-5750-022X
• Dionatan Schmidt UNIPAMPA https://orcid.org/0009-0006-1453-6860
• Aline Lunkes UNIPAMPA https://orcid.org/0000-0002-1846-4853
• Diego Kreutz UNIPAMPA https://orcid.org/0000-0003-0830-0238

DOI: https://doi.org/10.5753/eniac.2026.31375

Resumo

Quantization compresses small language models but erodes the long tail of rare knowledge. For a Brazilian deployment, that tail includes regional cultural factors, so a practitioner choosing a bit width needs to know whether any loss follows from the fact that it is Brazilian or merely from its being rare. We present CULTURAQUANT, a multiple-choice probe drawn from Wikidata that asks which region a Brazilian cultural figure or event originates from. Each item is paired with an equally rare cultural fact from another country, so the comparison isolates Brazilian origin rather than rarity. Across 11 open models at 8-bit and 4-bit, measuring the 6 that answer above chance, the Brazilian items erode no faster at 8-bit: the pooled differential is +0.2 percentage points (pp) and every per-model interval crosses zero, excluding gaps beyond about 1.8 pp. At 4-bit, a penalty appears only in the Qwen2.5 family. The bit-width budget safe for rare facts is also safe for rare Brazilian facts.

Referências

Almeida, T. S. et al. (2023). BLUEX: A benchmark based on brazilian leading universities entrance exams. In BRACIS.

Almeida, T. S. et al. (2025a). Building high-quality datasets for Portuguese LLMs: From common crawl snapshots to industrial-grade corpora. arXiv:2509.08824.

Almeida, T. S. et al. (2025b). PoETa v2: Toward more robust evaluation of LLMs in portuguese. arXiv:2511.17808.

Almeida, T. S. et al. (2025c). TiEBe: Tracking language model recall of notable worldwide events through time. arXiv:2501.07482.

Azime, I. A. et al. (2025). ProverbEval: Exploring LLM evaluation challenges for low-resource language understanding. In Findings of NAACL.

Badhe, S. et al. (2026). Long-tail knowledge in large language models: Taxonomy, mechanisms, interventions and implications. arXiv:2602.16201.

Borgersen, K. A. and Goodwin, M. (2025). English k-quantization of LLMs does not disproportionately diminish multilingual performance. arXiv:2503.03592.

Chang, H. et al. (2024). How do large language models acquire factual knowledge during pretraining? In NeurIPS.

Chimoto, E. A. et al. (2026). Calibrating beyond english: Language diversity for better quantized multilingual LLMs. In EACL.

Chiu, Y. Y. et al. (2025). CulturalBench: A robust, diverse and challenging benchmark for measuring LMs’ cultural knowledge through human-ai red-teaming. In ACL.

Corrêa, N. K. et al. (2025). Tucano: Advancing neural text generation for portuguese. Patterns.

Dettmers, T. et al. (2022). LLM.int8(): 8-bit matrix multiplication for transformers at scale. In NeurIPS.

Ebrahimi, A. et al. (2022). AmericasNLI: Evaluating zero-shot natural language understanding of pretrained multilingual models in truly low-resource languages. In ACL.

He, Y. et al. (2025). Supposedly equivalent facts that aren’t? entity frequency in pre-training induces asymmetry in LLMs. In COLM.

Hooker, S. et al. (2019). What do compressed deep neural networks forget? arXiv:1911.05248.

Hooker, S. et al. (2020). Characterising bias in compressed models. arXiv:2010.03058. Jodas, D. S. et al. (2024). Impact of quantization on large language models for Portuguese classification tasks. In CIARP.

Joshi, P. et al. (2020). The state and fate of linguistic diversity and inclusion in the NLP world. In ACL.

Kapelinski, C. and Kreutz, D. (2026). Not all 4-bit quantizers are equal: Deployment-time mitigation of PII leakage in fine-tuned small language models. In SBSeg.

Kwon, W. et al. (2023). Efficient memory management for large language model serving with PagedAttention. In SOSP.

Lee, J. et al. (2025). Exploring the trade-offs: Quantization methods, task difficulty, and model size in large language models from edge to giant. In IJCAI.

Marchisio, K. et al. (2024). How does quantization affect multilingual LLMs? In Findings of EMNLP.

Mekala, A. et al. (2025). Does quantization affect models’ performance on long-context tasks? In EMNLP.

Nguyen, C. V. et al. (2025). A survey on small language models. In RANLP. Pinhanez, C. et al. (2024). Harnessing the power of artificial intelligence to vitalize endangered indigenous languages: Technologies and experiences. arXiv:2407.12620.

Rath, P. K. and Maliakkal, R. (2026). Quantization undoes alignment: Bias emergence in compressed LLMs across models and precision levels. arXiv:2605.15208.

Santos, M. L. O. and Campelo, C. E. C. (2023). Benchmarking quantized LLaMa-based models on the brazilian secondary school exam. In CBIC.

Silva, G. et al. (2026). Lost in quantization: Activation outliers explain language-specific FP8 sensitivity in Llama-3. In PROPOR.

Singh, S. et al. (2025). Global MMLU: Understanding and addressing cultural and linguistic biases in multilingual evaluation. In ACL.

Tamiosso, G. L. et al. (2026). Evaluating small language models for english-to-portuguese translation: Impact of model scale and quantization. In PROPOR.

Tanwar, E. et al. (2025). Do you know about my nation? investigating multilingual language models’ cultural literacy through factual knowledge. In EMNLP.

Vieira, I. et al. (2026). ALBA: A european portuguese benchmark for evaluating language and linguistic dimensions in generative LLMs. In PROPOR.

Wang, J. et al. (2025a). SLMQuant: Benchmarking small language model quantization for practical deployment. In ACM Multimedia Workshops.

Wang, Q. et al. (2025b). Through a compressed lens: Investigating the impact of quantization on factual knowledge recall. arXiv:2505.13963.

Zhou, C. et al. (2025). Task-stratified knowledge scaling laws for post-training quantized large language models. arXiv:2508.18609.