AnalysisAI ModelsNovember 11, 2026

Blog: a linear program can replicate GPT-6 Astra's weight edits

Read original source →nathan.rs

A blog post argues Astra's weight-space edit of Qwen3.5-0.8B — which lifted MATH-500 from 7.9% to 48.1% by changing 0.016% of weights — was mostly an LP solved by trial and error over ~8 hours and 28 forward passes. Two ~1-minute LP solves match it with 8x fewer changed weights.

1 source

More stories today

Open the live feed