Researchers question whether transformer attention needs separate Q, K, V projections

A systematic study challenges a foundational assumption in transformer architecture: whether query, key, and value projections are all necessary. The findings could simplify model design and potentially reduce training overhead—relevant for indie developers building or fine-tuning language models on constrained budgets.