AnalysisAI ModelsNovember 4, 2026

MIT paper: GPT-6 Astra edits Qwen weights, lifts MATH-500 to 48.1%

Read original source →huggingface.co

MIT CSAIL researchers report GPT-6 Astra natively read and wrote Qwen3.5-0.8B's weights, changing just 0.016% of them to lift MATH-500 from 7.9% to 48.1% with no gradients or training. The authors say the edit mostly removed a reasoning-loop failure mode, and most experiments beyond the Qwen one were small.

People · Omar Khattab, Alex Zhang

How this story unfolded

2 days · 5 reports · 13 community posts · 18 of 22 shown

  1. Nov 2
  2. Nov 3
  3. Nov 4

More stories today

Open the live feed