Paper reveals LM Head as gradient bottleneck amid Anthropic vocab size rumors

JFPuget · x · 2026-08-15

Addressing rumors about Anthropic's model vocabulary size, this post cites the paper "Lost in Backpropagation: The LM Head is a Gradient Bottleneck". The paper argues that projecting output features to the vocabulary dimension creates not just an expressivity bottleneck, but an optimization bottleneck. Backpropagating V-dimensional gradients through a rank-D linear layer causes unavoidable compression, altering training feedback for most parameters. Empirical measurements show 95-99% of gradient norm is suppressed by the output layer, leading to suboptimal update directions.

Original post →

More from Models

Models channel →