PlurVA-LLM-2026 Shared Task Track-1: Pluralistic Value Alignment in LLMs via Multilingual Fine-Tuning and Threshold Calibration

2026-09-26

AI Summary

This study addresses the challenge of value alignment for large language models in low-resource languages, specifically Chinese, Indonesian, and Sinhala. Building upon Llama 3.1, we propose a multilingual alignment framework tailored for resource-constrained scenarios. Methodologically, we design language-specific data augmentation strategies encompassing option permutation, voting-based expansion, and binary reconstruction, integrated with QLoRA-efficient fine-tuning and an expanded SinhalaMMLU benchmark. Furthermore, a conditional threshold calibration mechanism is introduced to optimize decision boundaries. Experimental results demonstrate that the proposed approach achieves accuracies of 0.785, 0.715, and 0.916 across the three languages, respectively, yielding a macro-average accuracy of 0.805. These findings indicate that our framework effectively enables precise cross-cultural alignment of diverse values under limited computational and linguistic resources.

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.

Problem

• Pluralistic Value Alignment
• Large Language Models
• Multilingual
• Cross-cultural
• Shared Task

Innovation

• Pluralistic Value Alignment
• QLoRA Fine-Tuning
• Multilingual Data Augmentation
• Conditional Threshold Calibration
• LLMs