Answering in the User's Variety: Calibrated Accommodation Pairs for Post-Training
Uniform Standard English replies can hide unequal treatment of users’ dialects during training.
Across 10 open model lineages, preference optimization favored matching British and Australian spelling over African American English (AAE) and Singlish, even after accounting for each reply’s prior likelihood. The authors introduce calibrated accommodation pairs, which reward matching the user’s dialect density. On one fully open model, this improved matching on untrained features without detectably increasing dialect use for Standard English users or reducing instruction following. One-sided preference pairs achieved similar gains only by overshooting the user’s dialect use.
HF Daily Papers' note
Across 10 open model lineages, preference optimization favored matching British and Australian spelling over African American English (AAE) and Singlish, even after accounting for each reply’s prior likelihood. The authors introduce calibrated accommodation pairs, which reward matching the user’s dialect density. On one fully open model, this improved matching on untrained features without detectably increasing dialect use for Standard English users or reducing instruction following. One-sided preference pairs achieved similar gains only by overshooting the user’s dialect use.
HF Daily Papers' note
score 4