Answers are grounded in this paper and cite their sources.
Large language models (LLMs) often produce repetitive outputs, limiting linguistic diversity in lower-resource languages. This study introduces KAZ-Div, a model-agnostic inference-time framework combining multi-candidate generation, task-specific validity filtering, semantic redundancy management, and cluster-based response selection. We evaluated four LLMs across three Kazakh-language tasks: paraphrase generation, grounded question generation, and synthetic intent-data generation. Evaluation included lexical and embedding-based diversity metrics, automated task-specific constraint checks, stage-wise ablation, alternative prompting and decoding baselines, independent embedding-based fidelity validation, parameter-sensitivity analysis, and downstream intent classification. Compared with conventional single-response generation, KAZ-Div generally increased lexical diversity; for paraphrasing, Distinct-2 increased from 0.322 to 0.719 for GPT-5.6 Sol and from 0.559 to 0.840 for Claude Sonnet 4.6. However, embedding-based semantic-diversity and source-fidelity effects varied across models. Same-pool ablation revealed no consistent lexical-diversity advantage over random candidate selection, indicating that multi-candidate generation accounted for much of the observed improvement. Independent evaluations using LaBSE and multilingual-e5-base supported the principal source-fidelity patterns but did not establish an additional fidelity benefit from semantic curation. The current experiments do not demonstrate a general downstream benefit of KAZ-Div: after matching the number of unique training examples, the differences for GPT-5.6 Sol and Claude Sonnet 4.6 were not statistically significant, and multi-seed XLM-RoBERTa effects varied in direction across generators. Grammaticality, naturalness, task validity, and syntactic variation were not independently assessed. Overall, KAZ-Div provides a structured framework for quality-constrained candidate management, whose practical value depends on the generator, task-specific requirements, and computational budget.
No takes yet. Share an insight, caveat, or question.
Tleubayeva et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: