Internal Representation Statistics Fail as Difficulty Signals for African-Language NLI
August 20, 2026
An analysis of 15 African languages reveals that internal representation statistics cannot reliably provide example-level difficulty signals for adaptive inference. The study also identifies significant data contamination in AfriXNLI, where English configurations share 1,047 of 1,050 examples with XNLI evaluation data.
HOW THIS AFFECTS YOU
●
researcherYou should avoid using representation statistics for difficulty estimation in these multilingual settings and be wary of XNLI data leakage.