AnalysisAI ModelsNovember 2, 2026

GPT-6 Astra edits Qwen3.5-0.8B weights to lift MATH-500 from 7.9% to 48.1%

Read original source →rikiyat.github.io

Astra changed just 0.016% of Qwen3.5-0.8B's weights, with no gradients or training loops, fixing reasoning loops Qwen's own model card warns about. IFEval tripled as a side effect; Astra also wrote small game-playing networks from scratch that outplay PPO-trained nets.

7 sources

More stories today

Open the live feed