Researchers mitigate LLM dialect bias via activation steering

A study of six instruction-tuned large language models reveals that state-of-the-art systems systematically rewrite African American English (AAE) into Standard American English. The authors present an end-to-end framework to audit and mitigate this bias, introducing conditional Dialect Group Invariance (cDGI) for isolation and feature-level localization analysis.

• Syntactic constructions, particularly negative concord, are identified as universal triggers for bias across all tested models.
• The researchers apply activation steering, a training-free method that extracts dialect directions via causal tracing to inject into bias-relevant layers.
• This approach reduces bias 5 to 20 times more effectively than prompting while preserving fluency in Standard American English.
• The work includes REAL-AAE, the largest real-AAE parallel corpus with 17,479 triplets from natural tweets.

The study provides a method to prevent LLMs from silently correcting dialects and releases a significant new resource for evaluating dialect bias.