XCOPA

New method boosts multilingual LLM performance without retraining

Researchers have developed a novel inference-time method to enhance the performance of multilingual large language models across different languages. This technique, called SAE-Based Steering, utilizes pre-trained spars…