Reproduce “Analyzing Dialectical Biases in LLMs for Knowledge and Reasoning Benchmarks” on OpenResearch

About this paper

Eileen Pan, Anna Seo Gyeong Choi, Maartje ter Hoeve, Skyler Seto, Allison Koenecke

Researchers at Cornell University and Apple quantified performance biases in large language models when encountering non-Standard American English dialects in knowledge and reasoning benchmarks, observing accuracy reductions up to 20% and identifying specific grammatical rules as primary drivers of these disparities.

Key insights

• LLMs consistently exhibit a "quality-of-service gap," performing correctly on Standard American English questions but failing on dialectal variants, with significant accuracy reductions (up to 20%).
• The extent of performance degradation varies by dialect, model, and dataset, with Singaporean English and African American English showing some of the largest drops, correlating with increased perplexity.
• Specific grammatical rules, including "existential it," "drop copula be NP," and "y’all," are disproportionately responsible for dialectal underperformance, accounting for 64-85% of degradation in certain American dialects.

Summary by alphaXiv.