[P] Structured Prompting for Extremely Low-Resource Languages: 80% → 5% Vocabulary Contamination, No Fine-Tuning

A research team tackled the challenge of making large language models work for extremely low-resource languages like Tulu, a Dravidian language with only 2 million speakers and no standardized script. Instead of fine-tuning models—which requires data that doesn't exist—they developed a novel 5-layer structured prompting technique. Their approach includes phonological grounding (injecting Tulu's retroflex consonants), morphological rules (contrasting with similar Kannada), negative constraints (suppressing Kannada bleed), Romanization standardization, and self-play synthetic examples. The results were dramatic: vocabulary contamination dropped from 80% to just 5% across multiple models including GPT-4o, Gemini 2.0 Flash, and Llama 3.1 70B, with native speaker validation showing 85% grammatical accuracy.

What's particularly interesting is that the negative constraint layer—explicitly telling models what NOT to generate—proved more effective than expected, raising questions about whether models are truly learning Tulu grammar or performing sophisticated constrained generation. The self-play loop for creating synthetic examples was surprisingly sensitive to critique prompt wording, revealing a bootstrapping challenge: you need to specify "correct Tulu" to a model that doesn't know it. This research opens new possibilities for preserving linguistic diversity without massive datasets, though questions remain about how far prompting can go before fine-tuning becomes necessary and whether this approach generalizes to other language pairs like Maithili/Hindi or Scots/English.