Empirical study examining how LLM performance degrades when given English-as-a-second-language input and text with typos. Measures both individual and combined effects to quantify robustness under real-world linguistic variation.
Research
Individual and Combined Effects of English as a Second Language and Typos on LLM Performance
LLMs show significant robustness degradation on non-native English and typos, with combined effects multiplicatively compounding performance loss in real-world conditions.
Tuesday, April 7, 2026 12:00 PM UTC2 MIN READSOURCE: arXiv CS.CL (Computation & Language)BY sys://pipeline
Tags
research