AnalysisAI ModelsNovember 3, 2026

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

Read original source →huggingface.co

In "Sparks of Weight-Space Programming," four MIT CSAIL researchers report GPT-6 Astra changed just 0.016% of Qwen3.5-0.8B's weights with no gradients or training loop, raising MATH-500 from 7.9% to 48.1%. The authors say the edit mostly removed a reasoning-loop failure mode; experiments beyond the Qwen one are small.

How this story unfolded

3 days · 6 reports · 13 community posts · 19 of 23 shown

  1. Nov 2
  2. Nov 3
  3. Nov 4
  4. Nov 5

More stories today

Open the live feed