Analyze why large language models experience zero-shot collapse on low-resource tasks and how few-shot scaling or fine-tuned transformers like BanglaBERT overcome complex political sentiment nuances.