arXiv 2610.28474 · cs.CV · 2026-10-20
Answering in the User's Variety: Calibrated Accommodation Pairs for Post-Training
Zhizhuo Kou, Heming Li, Jian Yang, Junyu Luo, Jingshu Peng, Sirui Han, Yike Guo
机构待核实
一句话要点
论文提出校准适应配对(CAP)方法,通过偏好数据让模型在保持标准英语的同时更准确匹配用户方言,发现传统训练方式会无意识偏向特定拼写,且统一回复掩盖了训练中的选择性偏见。
CAP enables dialect-aware responses while preserving Standard English, revealing hidden biases in uniform model outputs.
要点
- 现有聊天模型无论用户使用何种方言,都默认输出标准英语,导致无法真实反映用户语言习惯,掩盖了训练中的选择性偏见。
- 提出校准适应配对(CAP)偏好数据,通过匹配用户方言密度来引导模型在不牺牲标准英语能力的前提下更准确响应用户方言。
- 在10个开源模型线路上实验表明,CAP显著提升了对非标准语言特征的响应能力,且未增加标准英语用户的方言使用或影响指令遵循性能。
- 意义与局限:CAP揭示了统一输出可能掩盖训练中的偏见,传统偏好数据无法有效捕捉方言适应性,需结合输出与更新过程进行审计。
English overview
Chat models often ignore user dialects, producing uniform Standard English responses. This paper shows that post-training stages introduce hidden biases favoring certain spellings like British or Australian over AAE or Singlish. The authors propose Calibrated Accommodation Pairs (CAP), a preference data method that rewards dialect matching without compromising standard English for standard users. Experiments across 10 model lines show CAP improves adaptation to non-standard features while preserving instruction-following and avoiding unintended dialect imposition.
只根据 arXiv 官方摘要,由开源模型整理,经数字/专名自动核对(必须在摘要中出现),以论文原文为准。
阅读 arXiv 原文 PDF Hugging Face 讨论 代码仓库
主题与学科
同主题:对齐
- Memadapter: Counterfactual Adaptation Against Memory-induced Sycophancy ▲62
- onPanda: Efficient Annotation of On-Policy Alignment Data for LLMs and Agents via Token-Level Correction ▲55
- Source Preference in the Wild: How LLM Agents Favor Items by Source, and How to Reduce It ▲47
- Diffusion Reward Models ▲35
同主题:强化学习
- False Frontiers: Diagnosing and Mitigating Co-Cheating in Self-Evolving Search Agents 2026-09-30 · ▲642
- The Teacher Is a Direction, Not a Destination: Extrapolating RL-Induced Representation Residuals in On-Policy Distillation 2026-09-29 · ▲511
- Player-1: Learning to Play from Synthetic Experience 2026-10-14 · ▲347
- Scaling Properties of Same-Family On-Policy Distillation 2026-09-26 · ▲327
- DiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Models 2026-10-08 · ▲325
数据说明
元数据来自 arXiv API(已核实);点赞来自 Hugging Face 每日论文;引用来自 Semantic Scholar。最近刷新 2026-10-20。
要点由本站根据摘要自写,不转载摘要原文,请阅读原文。