Meta, NYU and Yann LeCun argue BERT-style encoders break under scaling

pbaylies · x · 2026-07-25

The shared paper argues that standard BERT-style encoders have a scaling bug: as pretraining compute grows, their frozen representations can get worse rather than better.

To fix this, the authors propose CrossBERT, a two-part design that separates representation learning from token reconstruction. The paper says this enables larger masking ratios and gradient collection across all tokens, improves throughput and sample efficiency, and yields monotonic scaling on MTBE (eng, v2) and frozen GLUE benchmarks.

Original post →

More from Research

Research channel →