Why Can't the MiniMax LLM Say "Ma Jiaqi"? Internal Investigation of Sparse Token Forgetting

MiniMax presents an internal investigation into a sparse token forgetting phenomenon observed in its widely used M2 model series.