MIT uses GPT-6 Astra to edit Qwen weights, lifting MATH-500 to 48.1%

On November 12, 2026, the AI section of DEV Community reported, citing VentureBeat, on an experiment conducted by MIT researchers. They used OpenAI's GPT-6 Astra to inspect and modify roughly 0.016% of the parameters in Qwen3.5-0.8B, lifting the model's MATH-500 score from 7.9% to 48.1%.

According to the report, the team attributed the model's originally low score primarily to repetitive "thinking" loops that consumed its output limit, rather than to a lack of mathematical reasoning ability. The intervention was concentrated on two tensors, and the experiment relied on tools, hypothesis testing, and validation, with runs executed under budgets of 10, 25 and 50 forward passes. Separately, developer Nathan Barry reported that a simpler linear-programming approach reproduced or surpassed much of the improvement achieved by Astra, while changing roughly an equivalent share of parameters.

The report also notes that, for developers maintaining narrow-purpose models, the experiment shifts the proposed maintenance workflow from retraining toward diagnosing and patching internal behavior, provided that human review is applied. Whether direct editing ultimately proves cheaper, more reliable, and more scalable than retraining remains an open question.