MIT CSAIL researchers reported on November 2 that GPT-6 Astra improved a small Qwen model by directly editing its weights. If you're following claims about AI improving AI, this experiment gives you a concrete repair to examine.
A model that finally finishes its answer can look suddenly smarter. A failure to stop repeating can hide what it could otherwise solve, so the size of a test-score change depends on the defect in the starting model as well as the repair. The researchers found a striking improvement, then included a second model that helps explain where the improvement came from.
What Astra changed
In the Weight-Space Programming experiments, Astra edited Qwen3.5-0.8B in a Python environment. It could run practice inputs, but couldn't train with gradients or see the held-out evaluation questions. Weights are the learned numbers that govern a network's behavior.
The successful run changed 131,072 weights, about 0.016% of the model. With 50 allowed forward passes, which run inputs through the network, MATH-500 rose from 7.9% to 48.1%. The researchers traced the change to a direction associated with repetitive reasoning.
The starting checkpoint is Qwen's released 0.8-billion-parameter model. This was an experiment on accessible model weights, using an agent with tools. Replicating it involves a model checkpoint and the researchers' editing setup.
Keeping the starting checkpoint and the exact edit lets another team separate two questions: whether the repaired model still performs well, and whether a fresh editing run finds the same repair. Those are different kinds of reproducibility. We'd like evidence for each.
The control result earns its space
The authors explain the successful fix plainly: "it lets the model finish answers it could already work out". They also tried Gemma 3 270M, which rarely suffered from the same looping problem. Astra's edits produced no established improvement of comparable size there.
The Qwen run located an obstacle to using an existing ability. The Gemma experiment asked for improvement without that convenient obstacle and found a much harder problem.
It also tells you what to ask when the next weight-editing result arrives. Was the starting model failing for one concentrated reason? Did the change help on problems the editing agent hadn't seen? A gain that survives those questions gives researchers something others can build on.
Why a targeted repair still matters
Weights are usually the output of an optimization process. Here, a general-purpose language model reasoned about what to change inside another network. Even with its narrow successful case, that is a different way to approach a model defect.
It fits the broader effort to automate parts of AI research. The useful evidence is the change an agent made and the held-out behavior that followed. Counting how long an agent ran would tell us much less.
The next convincing demonstration would improve a model whose answers already finish. The Gemma run gives researchers a concrete place to start.
Cover: AI-generated editorial illustration by TMRW using MIT's official monogram. The numerical sheets are illustrative placeholders, not the experiment's weights.
