125M 'decision' models matched 200x larger LLMs — prior-art spat shadows Jev

zmkzmkz · x · 2026-09-25

After Delip Rao noted typesafeai's Jev shares its abstraction with autorubric, researcher ashabrawy201 pointed to their 2024 Statement Tuning paper: 125M universal 'decision' models matching 200x larger LLMs on unseen tasks, with inference mapping to Jev's decision head — a prior-art debate around the small-model-plus-decision-head paradigm.

Original post →

More from Research

Research channel →