AI

Turns out that token embeddings in smaller language models collapse into a narrow cone as they pass through transformer layers which reduces their representation expressivity. And the effect is much more servere in smaller models than larger ones. So one of the reasons larger models outperform smaller ones is due to having better organization of latent representations. Good news is that you can use a training objective that spreads embeddings uniformly across the representation space to counter the problem which means smaller models could be a lot more capable.