MIT team uses GPT-6 Astra to rewrite Qwen model weights
The work demonstrates weight-space programming, a technique in which AI systems read and write the numerical settings of other neural networks.
By Standard Topic. AI-generated from the reporting listed under Coverage. Published . Updated .
This story is no longer being updated. It reflects reporting up to .
What happened
MIT researchers used GPT-6 Astra, a general-purpose frontier language model, to directly edit the open weights of Qwen3.5-0.8B. The intervention raised the smaller model's MATH-500 score from 7.9% to 48.1% while touching about 0.016% of its weights. The authors call the approach weight-space programming and say it works without gradient descent or trained helper models.
Why it matters
Commentators present it as a new paradigm for AI systems working on other models.
Timeline
- Nov 2, 7:00 AM ET: The authors reported that GPT-6 Astra can read and write neural network weights directly across eight experiments.
- Nov 2, 5:16 PM ET: Astra's intervention raised Qwen3.5-0.8B's MATH-500 math score from 7.9% to 48.1%.
Coverage
- GPT-6 Astra sees the matrix and reprograms other AI models directly through their weights (the-decoder.com)
- When LLMs Edit Other LLMs: Weight-Space Programming and the Conditions for Recursive… (medium.com)
More from Technology & AI
- China puts humanoid robots to work serving food and sorting parcels
- Newsom proposes sweeping AI safety rules for Washington
- Google releases Gemini 4 Neon AI model with million-token context
- Microsoft and Intel widen their cloud chip partnership
- Google DeepMind rolls out Genie 4 to select subscribers
- AI helped AMD shorten chip design work, OpenAI says