MIT researchers report that Astra repaired Qwen's reasoning loops by editing weights
MIT CSAIL researchers say GPT-6 Astra edited 0.016% of Qwen3.5-0.8B's weights, raising MATH-500 from 7.9% to 48.1% in their tests. It found the patch using practice problems, 50 forward passes and direct weight edits without gradient descent. The gain came from reducing repetition; in a separate Gemma 3 270M test, the score went from 6.8% to 7.0% with the same forward-pass budget.
ThereforeResearchers can test the repair themselves using the released experiment logs and edited Qwen weights.
rikiyat.github.io research post · VentureBeat experiment analysis