Tag
This paper presents the first large-scale empirical analysis of 118 transformer models, revealing critical performance walls where success rates drop from 88.1% at 512 tokens to 0% at 2048 tokens, challenging prevailing scaling assumptions.