Pluralistic value alignment asks whether a language model can adapt its judgements to distinct, locally grounded value systems rather than reflecting a single aggregated notion of human preference. The PlurVA-LLM shared task evaluates this setting by unifying three independently constructed, culturally grounded multiple-choice value dilemmas across Chinese, Indonesian, and Sri Lankan datasets. We present a single-model submission based on Qwen3.5-4B with LoRA fine-tuning, trained only on the 1,359 labelled development instances provided by the task. Our method combines three ideas motivated directly by the task structure: soft-label supervision for Indonesian annotator-vote distributions, consistency regularisation across option permutations to reduce answer-position bias, and a staged multilingual curriculum with cross-validation-based iteration selection. On the hidden test set, our system achieves a macro-average accuracy of 0.8062, with 0.7788 on Chinese, 0.7139 on Indonesian, and 0.9260 on Sri Lankan, outperforming a manually tuned reference curriculum from our internal development. These results show that a single parameter-efficient model can unify culturally and structurally distinct value-alignment benchmarks under a low-resource training budget.

Figures of Speech (FOS) consist of multi-word phrases that are deeply intertwined with culture. While Neural Machine Translation (NMT) performs relatively well with the figurative expressions of high-resource languages, it often faces challenges when dealing with low-resource languages like Sinhala due to limited available data. To address this limitation, we introduce a corpus of 2,344 Sinhala figures of speech with cultural and cross-lingual annotations. We examine this dataset to classify the cultural origins of the figures of speech and to identify their cross-lingual equivalents. Additionally, we have developed a binary classifier to differentiate between two types of FOS in the dataset, achieving an accuracy rate of approximately 92%. We also evaluate the performance of existing LLMs on this dataset. Our findings reveal significant shortcomings in the current capabilities of LLMs, as these models often struggle to accurately convey idiomatic meanings. By making this dataset publicly available, we offer a crucial benchmark for future research in low-resource NLP and culturally aware machine translation.