MIT researchers report GPT-6 Astra editing Qwen's weights directly
1Nov 2 3:00 PM · 5d ago · 2 articles · 1 source · development 1 of 2
A study describes using GPT-6 Astra to manipulate Qwen's open weights without retraining, boosting its math performance by addressing repetitive reasoning patterns; the paper also notes Astra wrote and ran a Python program to perform the edit after roughly 50 trial-and-error attempts, and a similar edit attempted on Gemma to remove sycophancy was unsuccessful.
science.slashdot.org
“Like gene editing for neural networks”
study researchers
MIT researchersStudy authorsGPT-6 AstraOpenAI model used as the editing toolQwenAlibaba's open-weight model, the edit targetGemmaGoogle's open-weight model, a second edit target
The whole story articlespoststhe bright band is this development · numbered dots are the others · click one to jump