Answering in the User's Variety: Calibrated Accommodation Pairs for Post-Training

The paper argues that chat models tend to answer in Standard English regardless of the user's dialect, and that a uniform answer can hide whether post-training treats dialect speakers alike. The authors propose tracing how each post-training stage, from supervised fine-tuning (SFT) to preference optimization, affects dialect accommodation, and introduce calibrated accommodation pairs for post-training. The supplied text is truncated before reporting results or methods in detail.

Why it matters

Output audits alone may miss whether training stages treat dialect speakers equitably, so stage-level tracing could help developers locate and correct disparities.

Read with care

Source text is truncated mid-sentence, so the proposed method and any results are not described; source claims not independently checked.