Post-training standardizes dialects: DPO favors British spelling over AAE and Singlish across 10 lineages

Zhizhuo Kou · hf · 2026-10-22

HKUST researchers trace how each post-training stage shapes whether models match users' language variety.

• Post-training standardizes replies to every informal cue across 10 open model lineages
• After adjusting for each reply's prior likelihood, the DPO update still favors matching British/Australian spelling over AAE and Singlish — even though models can produce AAE features when asked
• They propose Calibrated Accommodation Pairs (CAP): preference data rewarding dialect-density matching while preserving Standard English for Standard-English users. On a fully open model, CAP raises convergence on untrained features with no dialect overshoot or instruction-following loss
• One-sided pairs only match the gain by imposing dialect on users
• Takeaway: uniform replies can hide selective updates; audits should read the update, not just the output

Original post →

More from Research

• Meta and MSU paper: LLM-built reasoning graphs are unreliable and their errors leak into confidence scores — rohanpaul_ai · 2026-10-22
• New paper: can first-price auctions efficiently allocate queries among strategic AI agents? — _onionesque · 2026-10-22
• RUC's LoopRec Brings Item-Level Adaptive Recursion with Early-Exit Gates to Sequential Recommendation — Zhanshuo Cao · 2026-10-22
• Anthropic's Agent4Sci: Frontier Agents Beat SOTA on 20 Scientific ML Benchmarks, but Only via Engineering Tricks — Anthropic · 2026-10-22
• NUS Paper Proposes PAPAV: A Unified Capability Framework from Computer-Use to Robot-Use Agents — showlab · 2026-10-22
• Cambridge's Rika Antonova on co-evolving robot bodies and brains, Co-design Gym expansion — contactrika · 2026-10-22

Research channel →