AnalysisAI ModelsNovember 17, 2026

Lattice of Latents adds per-layer recurrent state to Transformers

Read original source →shivamduggal4.github.io

A 406M-parameter Lattice model cut validation perplexity from 15.41 to 14.40 versus a parameter-matched Transformer on 20B FineWeb-Edu tokens. Combined with hyper-connections and input residuals, perplexity fell to 13.61.

1 source

More stories today

Open the live feed