Publications

Systematically evaluates prompt engineering strategies, including instruction language, few-shot scaling, and cultural framing, across diverse African languages and tasks. The study distinguishes between model limitations and prompt design issues, providing evidence-based guidelines for optimizing performance in low-resource classification settings.

Conducts a systematic scaling study on 16 African languages using AfriXNLI, results show that small evaluation sets can produce unstable, variably biased results. We find scaling behavior is often non-monotonic and language-specific, with high variance in low-resource regimes. Our work stresses the need for larger, more representative evaluation sets and stronger multilingual modeling for reliable benchmarking.