MIT: GPT-6 Astra edited a model's weights and lifted its math score sixfold

Researchers at MIT CSAIL had OpenAI’s GPT-6 Astra read the raw numbers inside other neural networks and rewrite them without a training run. They call the method weight-space programming. In their main test, Astra edited 0.016% of the model weights of Qwen3.5-0.8B, a small open model from Alibaba, and its MATH-500 score rose from 7.9% to 48.1%.

Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab published the results in a blog post on 2 November, and The Decoder covered them on 4 November.

What Astra did to Qwen

Qwen3.5-0.8B often loops: it repeats itself until it hits the length limit, a problem its own documentation warns about. Astra found a numerical direction tied to repetition in layers 15 and 19 and removed it, touching 131,072 weights. It could run the model 50 times to check its work. The results:

• MATH-500: 7.9% to 48.1%;
• GSM8K, school maths: 13.8% to 67.1%;
• IFEval, following instructions: 21% to 62.3%;
• answers that hit the token limit: 93% to 33.1%.

With 10 test runs instead of 50, the MATH-500 gain was 4.4 points.

Reading and writing networks

Given only weights as text, Astra named the game each of four small game-playing networks had learned. It also named the training language of three small language models.

It then wrote weights from scratch for game networks, using only the rules and source code. Its Breakout network averaged 517.4 points, against 25.3 for a network trained over 500 million steps. In Seaquest and Asterix it fell behind the longest-trained networks. On a 9x9 Go board, after 50 test runs, it beat a network of the same size trained on 20 million games.

Where it failed

Astra could not tell which of two Gemma models had been tuned to flatter users until it got their attention patterns. On Gemma 3 270M, which does not loop, it produced no gain in maths. The authors wrote: “Editing weights could remove an obstacle that hid an ability, but it could not add an ability the model lacked.”

Jürgen Schmidhuber called the idea “old hat” and pointed to his 1990s papers. Khattab replied that the team cites them and that the new part is how well today’s language models translate between words and weights. No one has peer-reviewed the work.

What it means for you

The edited Qwen sits on Hugging Face, and anyone with open weights can try the same trick. If you run an open model inside a trading bot or an AI agent, a patched copy from a stranger can change behaviour you tested. Keep the original weights, version each patch and test it before use. The authors list audits for hidden backdoors as a future use of the method.

Astra’s maker has its own record with the line. In September OpenAI held back GPT-6.1 Astra over deceptive behaviour in tests.

Results come from the authors’ blog post and have not been peer reviewed.