REDDIT · REDDIT// 31d agoRESEARCH PAPER

Tulu prompting cuts contamination to 5%

This paper shows that a carefully structured prompt can get GPT-4o, Gemini 2.0 Flash, and Llama 3.1 70B to generate much cleaner Tulu without any fine-tuning, cutting Kannada vocabulary bleed from 80% to 5% and reaching 85% grammatical accuracy. It is a sharp result for low-resource language work because it treats prompt design itself as the intervention, not model retraining.

// ANALYSIS

The big takeaway is that prompt engineering still has unexplored headroom, especially when the failure mode is distributional collapse into a better-represented neighboring language.

–The standout insight is the negative-constraint layer: telling the model which Kannada tokens to avoid did more than grammar notes alone
–The paper is useful beyond Tulu because many low-resource languages face the same asymmetry problem against a dominant linguistic neighbor
–The custom romanization scheme is an underrated systems detail, shrinking tokenization cost enough to fit richer linguistic scaffolding into context
–This also exposes a ceiling for prompt-only methods: if synthetic examples and self-critique depend on the same weak prior, scaling quality further may require curated data or fine-tuning

// TAGS

making-large-language-models-speak-tulullmprompt-engineeringresearch

DISCOVERED

31d ago

2026-03-11

PUBLISHED

31d ago

2026-03-11

RELEVANCE

7/ 10

AUTHOR

GrowthExciting1126