Story

On air. The free wire runs 15 minutes behind live. See pricing

World, crypto, finance and tech news, scored for impact.

TechHacker NewsFiled 35 h ago

Latent Lattice Language Models

Open the sourceReported by Hacker News, shivamduggal4.github.io, 17 November 2026

Not investment advice. Muhabbit scores how much a story may matter, not which way to trade.

MIT and Adobe researchers introduced the Lattice of Latents, a language-model architecture that adds per-layer horizontal latent streams persisting across tokens alongside the standard vertical residual stream, directly conditioning attention, KV-cache writes, MLP computation and decoding. In experiments on 20B FineWeb-Edu tokens, a 406M-parameter Lattice model cut validation perplexity to 14.40 from 15.41 for a parameter-matched Transformer, and combining the approach with hyper-connections and input residuals lowered perplexity further to 13.61. The work, by Shivam Duggal, Zongze Wu, William T. Freeman, Antonio Torralba, Richard Zhang and Eli Shechtman, is published with code promised soon; next steps include external replication and tests at larger parameter and token scales.

Muhabbit's AI wrote this summary in its own words from Hacker News's report. It can get details wrong, so check the source.

Named in this story

MIT, Adobe (ADBE), Eli Shechtman, Richard Zhang, Antonio Torralba, William T. Freeman, Zongze Wu, Shivam Duggal

A linked name opens that asset's sentiment page.