GPT-6 Astra fixed Qwen's looping: accuracy rose from 7.9% to 48.1%

A local model receives a math problem, starts reasoning and gets stuck repeating itself. Tokens are consumed, the user waits, and no completed answer arrives. A team at MIT CSAIL demonstrated an experimental way to fix this failure: ask another language model to locate the source of the loop and adjust the affected network's numerical parameters. For developers, this offers a useful lead: low accuracy sometimes stems from a model failing to finish its solution.

In the published experiment, GPT-6 Astra edited Qwen3.5-0.8B. After the intervention, its score on the MATH-500 mathematics benchmark rose from 7.9% to 48.1%. The authors call the approach weight-space programming—programming through a model's weights. The results and experimental conditions are detailed in the researchers' publication.

What exactly was fixed

Weights are the numbers that determine how a neural network processes input. They are normally changed during training. Here, the agent was given access to the parameters and the ability to test hypotheses, while conventional gradient-based fine-tuning was prohibited.

A simple example illustrates the problem: the model has already derived the required formula, then explains the previous step again and returns to it yet another time. The response cuts off when the token allowance runs out. During evaluation, that run may count as a failure even if the useful part of the solution has already appeared.

According to the researchers, the agent found a component in two Qwen layers that sustained repetition and reduced its contribution. The change affected just 0.016% of the parameters. This resembles troubleshooting a specific fault: first identify where the unwanted behavior originates, then test whether a small intervention helps.

The workflow itself is of practical interest. The agent analyzes the network's architecture, proposes an explanation for the failure, changes parameters and receives feedback. For a team maintaining its own model, this could be an additional diagnostic tool. The experiment does not yet provide a universal recipe for fixing every kind of repetition.

Why the accuracy gain needs caveats

The authors explicitly link the improvement to eliminating looping. The result therefore cannot automatically be interpreted as the acquisition of new mathematical knowledge. If the model previously failed to reach a final answer, removing that obstacle alone can sharply increase its score.

For a product, the distinction matters. Users need a completed result, and fixing repetition can be highly useful. But a product manager promising stronger mathematical reasoning in general after such an intervention needs to test that claim separately on the service's own tasks.

The study includes a revealing test on another small model, Gemma 3 270M. The researchers did not achieve a comparable improvement. This negative result limits the conclusion: successfully fixing one pronounced failure does not mean the agent has learned to reliably improve any network.

Nor does the absence of further training make the procedure free. Analysis, test runs and evaluation require computation. When comparing this approach with other fixes, teams will need to account for all those costs. The cost of getting it wrong matters too: even a small parameter change must be assessed across the model's overall behavior, including the tasks it was originally chosen to perform.

What to check on your own model

Start with logs of unsuccessful responses. How often does the model repeat itself? How many requests end at the limit? Do these cases differ from ordinary calculation or factual errors? Labeling them this way helps establish which problem actually needs solving.

Next, assemble a test set that the agent does not use when selecting changes. Use it to compare answer quality, generation length and wait times before and after the intervention. It is also useful to test routine workflows separately: improvement on one test may come with deterioration on another.

For teams working through a closed API, directly editing weights is usually unavailable. But the diagnostic lesson still applies: distinguish knowledge errors, reasoning errors and unfinished answers. That distinction determines the next step—changing the model, generation settings or the application's design. The MIT experiment provides grounds for examining the causes of failures more closely before budgeting for another training run.