Illustration, AI-generated from the story text. Tap a panel to read the story through it.

1 / 4MIT researchers had GPT-6 Astra edit a small Qwen model's weights directly, without training×

The edit

Researchers at MIT CSAIL, led by PhD student Rikiya Takehi with Omar Khattab, report that GPT-6 Astra improved Qwen3.5-0.8B's math by changing 0.016% of its weights by hand, with no gradient descent: it reasoned about what parts of the network compute and set new values.

The claims

They report MATH-500 scores rising as much as sixfold and IFEval about threefold, and Astra writing a Breakout-playing policy network from scratch. The authors pitch it as a possible route to fixing models without forgetting, and to self-improvement.

The skeptics

The skeptics have the stronger reading of the method. Matthew Honnibal points out that the headline Qwen result used Python and ran the model on up to 50 batches of practice problems, so it is not "just looking at the raw weights," and that the from-scratch Breakout network beat a baseline on Breakout but lost on Seaquest and Asterix. He still credits the Qwen result as real.

The limits

It is one small model and a few benchmarks, without peer review.

Sources and the full story

‹›

MIT researchers had GPT-6 Astra edit a small Qwen model's weights directly, without training

Researchers at MIT CSAIL, led by PhD student Rikiya Takehi with Omar Khattab, report that GPT-6 Astra improved Qwen3.5-0.8B's math by changing 0.016% of its weights by hand, with no gradient descent: it reasoned about what parts of the network compute and set new values. They report MATH-500 scores rising as much as sixfold and IFEval about threefold, and Astra writing a Breakout-playing policy network from scratch. The authors pitch it as a possible route to fixing models without forgetting, and to self-improvement.

The skeptics have the stronger reading of the method. Matthew Honnibal points out that the headline Qwen result used Python and ran the model on up to 50 batches of practice problems, so it is not "just looking at the raw weights," and that the from-scratch Breakout network beat a baseline on Breakout but lost on Seaquest and Asterix. He still credits the Qwen result as real. It is one small model and a few benchmarks, without peer review. SourcesABC

Sources

• AWeight-space programming (Rikiya Takehi, MIT CSAIL, not opened)rikiyat.github.io
• BForget retraining, AI can now tweak another AI's weights (The Register)theregister.com
• CAstra read-write claims are misleading (Matthew Honnibal)honnibal.dev

This story ran in the Morning Brief of November 3rd 2026.