Researchers have released the first survey-grounded set of resources for testing whether large language models handle Sri Lankan social values correctly, in a paper posted to arXiv on 22 July.
The suite, called LKValues, starts from a trilingual survey instrument in Sinhala, Tamil and English that combines items adapted from established international value frameworks with an LLM-assisted stage meant to surface constructs those global questionnaires miss. Responses from 205 people yielded 40 majority-endorsed Sri Lankan societal values.
Those values organise two published datasets. LKvaluesIT is a Sinhala–English instruction corpus of 150,000 scenario-based instances derived from news text. LKvaluesBench is a 1,000-instance evaluation set. Both are on GitHub.
The team evaluated a mix of proprietary and open-weight models against the benchmark, then fine-tuned three open-weight base models on the corpus: Qwen3.5-4B-Base, Qwen3.5-9B-Base and Aya-Expanse-8B-Base. Newer and larger models still showed both low-resource language and cultural value-alignment gaps, the authors report. Fine-tuning improved the Qwen-family models in English and Sinhala, reducing invalid outputs and narrowing the disparity between the two languages, though the paper states that gains were not uniform across model families.
The author list is the second notable element. Six of the ten authors, including the first author and corresponding author Deyi Xiong, are at the TJUNLP Lab in the School of Computer Science and Technology at Tianjin University. Sri Lankan institutions are represented by one researcher each at the University of Moratuwa and the University of Colombo, with the remaining authors at Massey University in New Zealand and Johns Hopkins University. Two of the three base models fine-tuned are Qwen.
The work is a preprint and has not been peer reviewed. A 205-person survey is a narrow base for a national values instrument, and the authors present the method as a replicable pipeline for low-resource, country-specific alignment rather than a settled account of Sri Lankan values. The survey instrument is trilingual, but the released datasets cover only Sinhala and English; Tamil, also an official language, does not appear in the published artefacts. The instruction corpus, being news-derived, carries the framing of the press that produced it.
Value alignment is emerging as a second layer of localisation work, distinct from the language-capability problem that has absorbed most low-resource NLP effort — a model can produce fluent Sinhala and still resolve a scenario on imported assumptions. Whether that layer matters commercially will show in procurement: whether any bank, telco or government service deploying Sinhala-language systems evaluates against LKvaluesBench.
Sources
- LKValues: Aligning Large Language Models with Sri Lankan Societal Values, arXiv:2607.20410, submitted 22 July 2026 (preprint, 37 pages).
- Project repository: github.com/NextME14/LKValues
