Could a specific flaw in an AI model be fixed without retraining it? A study by MIT CSAIL researchers offers one carefully bounded example.
What changed in the experiment?
The team asked GPT-6 Astra to inspect and modify Qwen3.5-0.8B, a small language model. It worked with the model’s weights: the numerical settings that shape a neural network’s behaviour. The researchers call the approach weight-space programming. Astra had Python access to the weights and could run the model on practice problems, but the edit did not use gradient-based training.
The model frequently became trapped in repeated reasoning and failed to finish its answers. In the experiment with a budget of 50 forward passes, Astra changed about 0.016% of its weights. The authors report that MATH-500 accuracy rose from 7.9% to 48.1%, and GSM8K from 13.8% to 67.1%. The share of MATH-500 responses hitting the length limit fell from 93% to 33%. These are results under the study’s settings, not a promise for other models.
What the result can and cannot show
There is a crucial distinction: removing a loop can let a model finish problems it could already solve. It does not necessarily add new mathematical ability. The researchers also tried Gemma 3 270M, which rarely looped, and did not establish a meaningful maths improvement. Humans defined the task, tools, budget and evaluation; this experiment does not demonstrate unrestricted autonomous self-improvement.
The work suggests a way to investigate and edit specific model behaviours. How reliably it transfers to other architectures remains a question for further independent testing.
