Fine tuning improves multilingual value alignment in large language models
PlurVA-LLM-2026 Shared Task Track-1: Pluralistic Value Alignment in LLMs via Multilingual Fine-Tuning and Threshold Calibration
Computation and Language
Summary
The paper tackles the challenge of aligning large language models with the diverse values of people from China, Indonesia, and Sri Lanka. The authors fine-tuned a language model called Llama 3.1 by using specific techniques tailored to each country's data, such as data augmentation and vote expansion. They also adjusted prediction thresholds to better fit the Sri Lankan context. Their methods improved the model’s accuracy in understanding different value perspectives across these languages.
What this means in practice
- •For multilingual ai developers: Enhance pretrained language models to respect cultural and regional value differences using selective fine-tuning and calibration techniques.
- •For chatbot designers: Create chatbots that better handle diverse value judgments by integrating pluralistic alignment methods across several languages.
Authors
Vihindi Kotalawala, Nevidu Jayatilleke
Abstract
We present our system for the PlurVA-LLM 2026 Shared Task Track-1, which focuses on pluralistic value alignment in the contexts of China, Indonesia, and Sri Lanka. For this resource-constrained track, we fine-tuned Llama 3.1 8B Instruct using 4-bit QLoRA. Our approach combines option-permutation augmentation for Chinese data, annotator vote expansion for Indonesian data, and binary reformulation with SinhalaMMLU augmentation for Sri Lankan data. We further applied conditional threshold calibration to the predictions for the Sri Lankan data. The final system achieved accuracies of 0.785 for Chinese, 0.715 for Indonesian, and 0.916 for Sri Lankan, resulting in an overall macro-average accuracy of 0.805.