作者位次 #8/10ACL 2026Findings· Findings of ACL
Reinforcement Learning with Semantic Rewards Enables Low-Resource Language Expansion without Alignment Tax
Zeli Su · Ziyin Zhang · Zhou Liu · Xuexian Song · Zhankai Xu · Longfei Zheng · Xiaolu Zhang · Rong Fu · Guixian Xu · Wentao Zhang
摘要
Extending large language models (LLMs) to low-resource languages often incurs an “align- ment tax”: improvements in the target lan- guage come at the cost of catastrophic forget- ting in general capabilities. We argue that this trade-off arises from the rigidity of supervised fine-tuning (SFT), w...