Chat models adapt answers to match user dialects with new calibration pairs
Answering in the User's Variety: Calibrated Accommodation Pairs for Post-Training
Computer Vision and Pattern Recognition
Summary
Chat models usually reply in standard English, ignoring how people actually talk in different dialects or slang. The authors show that during training, models often become more uniform and less sensitive to dialects like African American English or Singlish. They created a new training approach called calibrated accommodation pairs (CAP) that encourages models to match a user’s dialect style without losing clear standard English responses for others. This helps make chatbots sound more natural to a wider variety of users without causing confusion for standard English speakers.
What this means in practice
- •For conversational ai developers: Train chat models that adapt responses to match users’ dialects without degrading standard English interactions.
- •For customer service platform teams: Improve user experience by tailoring chatbot replies to the user’s language style while maintaining clear communication.
Authors
Zhizhuo Kou, Heming Li, Jian Yang, Junyu Luo, Jingshu Peng, Sirui Han, Yike Guo
Abstract
People adapt how they talk to the person in front of them. However, chat models tend to respond in Standard English regardless of the dialect used by the user. A uniform answer can hide whether training treats these users alike, and output audits cannot tell as well. We trace how each post-training stage, from supervised fine-tuning (SFT) to preference optimization (DPO) and reinforcement learning, changes a model's tendency to match the user, and read what each update prefers from two content-matched replies, one matching the user and one not. Across 10 open model lineages, post-training standardizes replies to every informal cue, yet after adjusting for how likely each reply already was, the DPO update still favours matching British and Australian spelling over AAE and Singlish on 10 lineages, and models still produce AAE features when asked. We then introduce calibrated accommodation pairs (CAP), preference data that reward matching users' dialect density while preserving Standard English for Standard-English users. On a fully open model, CAP raises convergence on features it was not trained on without a detectable increase in dialect for Standard-English users or a cost in instruction following. In contrast, one-sided pairs reach a similar gain only by overshooting the user and imposing dialect. Our findings show that uniform replies can hide selective updates and audits should read the update as well as the output. Code: https://github.com/kouzhizhuo/cap-accommodation.