publication

MIT-affiliated researchers publish GPT-6 Astra direct edits to Qwen weights without gradient training

Researchers Takehi, Teoh, Zhang and Khattab published controlled weight-space programming experiments on 2 November 2026. With Python access and up to 50 forward-pass experiments, GPT-6 Astra edited 131,072 Qwen3.5-0.8B MLP weights (0.016%) to reduce reasoning loops; authors report MATH-500 7.9% to 48.1% and GSM8K 13.8% to 67.1%. This is the authors' controlled experiment and published logs, not independent broad replication, weight self-modification of Astra, generalized intelligence explosion or demonstrated transfer beyond this setup.