People naturally adjust how they speak to match the person in front of them, but chat models do not. A new preprint on arXiv points out that these models tend to answer in Standard English even when a user writes in a different dialect. The authors suggest that this uniformity may hide whether the model is actually treating all users fairly.
The paper proposes a post-training technique called "calibrated accommodation pairs" to address this. The idea appears to be to teach models to match the user's variety of language. The abstract does not provide further details on how the pairs are constructed or evaluated.
The preprint highlights that dialectal diversity remains a challenge for language models. If models only ever respond in Standard English, they may not fully serve users who communicate in other dialects. The authors argue that a uniform answer can obscure differences in how users are treated.