AnalysisAI ModelsNovember 18, 2026

LLMs designing nets from scratch skip Transformers for sparse tables

Read original source →arxiv.org

POSTECH researchers gave Claude Opus 5.5, GPT-6.1 Sol and GPT-6 Astra ten labeled questions and a code sandbox with no gradient descent; all three built sparse lookup-table models ~1,800x smaller that beat same-sized Transformers. Pretrained LLMs stayed more accurate.

2 sources

More stories today

Open the live feed