MIT tests lift a small model’s maths score from 7.9% to 48.1%
MIT researchers used GPT-6 Astra to edit 0.016% of Qwen3.5-0.8B’s weights, lifting its MATH-500 score from 7.9% to 48.1%. The change removed a repetition failure that prevented the model finishing answers. A separate Gemma 3 270M test showed no established maths improvement.
Why it matters: The gains came from repairing a specific failure, which sets a concrete limit on claims about automated AI improvement.
rikiyat.github.io
AI Impact Hub