PlurVA-LLM-2026 Shared Task Track-1 and 2: Preference Learning for Pluralistic Value Alignment
Abstract
AbstractWe describe the Modgenix submissions to both tracks of the PlurVA-LLM 2026 Shared Task on pluralistic value alignment across China, Indonesia, and Sri Lanka. For the resource-constrained track, we use one dense 7.94B-parameter Gemma 4 E4B model for all three datasets. We recast each option as a binary candidate judgment and train rank-8 LoRA parameters using classification and pair- wise preference losses. Deterministic augmen- tation permutes answer positions, balances lan- guages, and creates a small Chinese core- action view. For the open track, we addition- ally evaluate a single GPT-5.4 model with ver- sioned multilingual prompts; its best macro accuracy is 0.795. The compact preference- trained system is stronger, so we submit it un- changed to both tracks. It obtains accuracies of 0.788, 0.683, and 0.954 on the Chinese, In- donesian, and Sri Lankan hidden sets, respec- tively, for a macro-average of 0.808 and first place on both final leaderboards. Error anal- ysis associates candidate preference learning with fewer downside-salience errors, while an objective mismatch between Indonesian anno- tator votes and consensus evaluation remains a key limitation.