Weak-to-strong (W2S) generalization is a promising framework for scalable oversight, yet existing evaluations often test students under matched train-test distributions. Therefore, we study W2S preference learning under zero-shot distribution shift and find that strong students trained on weak preference labels can appear successful in-distribution while failing to transfer across preference datasets. We provide evidence consistent with a representational failure mode in which weak-supervised fine-tuning may move the strong model toward source-domain features instead of maintaining broadly transferable preference representations. To mitigate this, we propose *Representation Anchoring* (Anchor), a simple yet effective regularizer that constrains excessive drift from the pretrained strong model’s representation space during fine-tuning, while still allowing task-relevant adaptation. Across preference domains, datasets, and model families, Anchor improves out-of-distribution transfer while maintaining competitive in-distribution performance. Together, our evaluation protocol, transfer-aware metrics, and method expose hidden brittleness in current W2S reward modeling and provide a practical path toward more robust preference transfer.

The Bahnar people, one of Vietnam’s ethnic minorities, represent an underserved community with limited access to modern technologies. Developing an effective Bahnaric-Vietnamese translation system is essential for fostering linguistic exchange, preserving cultural heritage, and empowering local communities by bridging communication barriers. With advancements in Artificial Intelligence (AI), Neural Machine Translation (NMT) has achieved remarkable success across various language pairs. However, the low-resource nature of Bahnaric, characterized by data scarcity, vocabulary constraints, and the lack of parallel corpora, poses significant challenges to building an accurate and efficient translation system. To address these challenges, we propose a novel hybrid architecture for Bahnaric-Vietnamese translation, with BARTBahnar as its core language model. BARTBahnar is developed by continually training a pre-trained Vietnamese model, BARTPho, on augmented monolingual Bahnaric data, followed by fine-tuning on bilingual datasets. This transfer learning approach reduces training costs while effectively capturing linguistic similarities between the two languages. Additionally, we implement advanced data augmentation techniques to enrich and diversify training data, further enhancing BARTBahnar’s robustness and translation accuracy. Beyond leveraging the language model, our hybrid system integrates rule-based and statistical methods to improve translation quality. Experimental results show substantial improvements on bilingual Bahnaric-Vietnamese datasets, validating the effectiveness of our approach for low-resource translation. To support further research, we open-source our code and related materials at https://github.com/ura-hcmut/BARTBahnar.