MIT CSAIL researchers reported on November 2 that GPT-6 Astra raised Qwen3.5-0.8B's MATH-500 score from 7.9% to 48.1% by editing its weights. Their experiment points to a specific repair: reducing repetitive reasoning loops. It doesn't establish a general method for adding mathematical ability. MrKeyoor has not reproduced the results.
The authors, Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab, call the approach weight-space programming. Astra used Python to inspect and alter the model, with practice problems and a budget of 50 forward passes, but no gradient descent. In that setup, a forward pass could process a batch of problems, so 50 passes should not be read as 50 individual answers. The experimental description separates practice data from the evaluation sets.
A small edit with a testable explanation
According to the released checkpoint's model card, Astra changed 131,072 weights, or 0.016% of the model, in layers 15 and 19. The card reports a separate rerun scoring 51.0% on all 500 MATH-500 problems, with four samples per problem. Its base-model check covered only 25 problems. That smaller baseline limits what readers can infer from the rerun's before-and-after comparison.
The logs repository supplies prompts, generated code and evaluation records, plus exact weight diffs and a script to apply them. It also says earlier prompt trials are omitted. These records make the reported edit available for inspection, while leaving the history of experimental choices incomplete. The repository identifies Astra as OpenAI's gpt-6-astra.
The failed follow-up matters
The researchers also tried Gemma 3 270M, which rarely looped. They found no established math improvement. That negative result helps define the contribution: removing a repetitive failure can let a model finish answers without demonstrating that it learned new reasoning skills.
For developers considering the published checkpoint, the next useful test is a matched evaluation of both versions on the same full dataset and serving settings. Check answer accuracy alongside how often generation exhausts its token budget. The model card records the rebuild and evaluation configuration, giving independent testers a concrete starting point.