Pinetree at SemEval-2026 Task 7: A Large-Scale Failure Analysis of Cultural Grounding in Language Models

AbstractUsing a simple prompting strategy without fine-tuning or retrieval augmentation, our system achieved 88.85% micro-average and 90.55% macro-average accuracy, ranking #4 overall on SemEval-2026 Task 7. Our primary contribution is a failure analysis of 5,241 incorrect predictions (11.15% of the dataset), categorized using the six-topic BLEnD taxonomy. Errors concentrate in Food (39.42%) and Holidays/Celebration/Leisure (15.76%), but within-topic error rates are highest on Family (21.04%) and Work life (20.45%), which topics with limited representational density. Global-brand attractor errors account for only 2.50% of failures and are tightly localized: 98.5% fall on a single template (most popular sport team) in four low-resource cultures. Outside these templates, brand-default effects are statistically negligible. These findings support representational sparsity and knowledge-density asymmetry, not ideological skew, as the dominant cause of cultural misalignment in everyday behavioral tasks.