The result: post-norm outperforms pre-norm across the board. Up to 2× better acceptance under template perturbation, 1.18× on long-context, 1.1× on standard benchmarks with 30% lower training cost.…