Reinforcement Learning Elicits Contextual Learning of Unseen Language Translation
Researchers have developed a reinforcement learning (RL) method to improve large language models' (LLMs) ability to translate unseen languages. This approach trains LLMs to extract and utilize linguistic information from provided context, rather than simply memorizing specific languages. The RL models, rewarded by a translation metric, demonstrated better performance on completely new languages compared to traditional in-context learning or supervised fine-tuning. AI
IMPACT Enhances LLM capabilities for low-resource language translation, potentially broadening access to information and communication.