GPT-6 Astra increases a small model score from 7.9% to 48.1%

The authors of a report tell that GPT-6 Astra can read and write the weights of a neural network. Weights are the numbers that set what a model does. Astra found a fault in a small model, Qwen3.5-0.8B, that causes reasoning loops. It changed 0.016% of the weights. The score on the MATH-500 test went from 7.9% to 48.1%.

How it works

Astra looks at the raw weights and finds what each part of the network does. It then changes the weights by reasoning. It does not write a training loop, which is the code that adjusts a model in small steps. The authors call this Weight-Space Programming.

The numbers

The IFEval test checks if a model follows instructions. Its score went up by 3 times. The authors also tell that Astra wrote small game networks from scratch. These networks beat networks that had 100 million steps of PPO training.

What is not known

The source gives no test by other teams. It shows one small model. The authors tell that large language models will change weights in the future. This is a prediction and not a result.

Insight: The skill that repairs weights can also hide changes

Astra changes 0.016% of the weights of a model to repair a fault. In October 2026, Stealthium told that a bad Python library can change weights in GPU memory while no file on disk changes. Models can read weights. The same skill looks like a help for repair and for attack, and the better reader has the control. Tell your vendor to show that its tools see weight changes.