Researchers at DeepSeek
Their state-of-the-art Global LBL load balancing loss function was outperformed by ShinkaEvolve's newly discovered loss function for Mixture-of-Experts model training
How media typically covers DeepSeek team
Research or work cited
Sakana AI introduces ShinkaEvolve, an open-source evolutionary framework that discovers new algorithms with LLMs and achieves unprecedented sample efficiency, solving Circle Packing with 150 samples versus thousands required by prior approaches.
“Their state-of-the-art Global LBL load balancing loss function was outperformed by ShinkaEvolve's newly discovered loss function for Mixture-of-Experts model training”