GPT-6 Astra edited a small model's weights to stop its loops, MIT team says

TL;DR

• MIT CSAIL researchers say GPT-6 Astra edited 0.016% of Qwen3.5-0.8B's weights and lifted its MATH-500 score from 7.9% to 48.1%, with no training.
• The gain came from removing a looping fault, and Astra worked in Python with 50 forward passes of the model.
• Nobody has peer reviewed the post, Matthew Honnibal says it mixes two different methods, and Sean Goedecke says a skilled human researcher could do the same tasks.