In a November 2, 2026, blog post, four researchers affiliated with MIT’s Computer Science and Artificial Intelligence Laboratory reported that GPT-6 Astra edited 131,072 weights in Alibaba’s Qwen3.5-0.8B, or 0.016% of the model. In the 50-forward-pass condition, they reported a MATH-500 score of 48.1%, up from 7.9%. The researchers said the edit targeted repetitive reasoning loops that kept Qwen from finishing answers. Their research post describes the method and results.
How the Qwen experiment worked
The researchers called the approach weight-space programming: using a language model to read or change neural-network weights for a specified task, without using gradient descent to calculate the edits. For the Qwen experiment, Astra used a Python kernel to work with the weights. It had no GPU, and the model ran with its thinking mode enabled.
Astra could choose practice problems from 7,469 GSM8K and 7,275 MATH training examples. The researchers allowed 10, 25 or 50 forward passes. Each pass could process up to 512 problems or read up to 1,024 texts. Astra did not receive the evaluation sets.
The edit changed 64 MLP output columns in each of layers 15 and 19. That amounted to 131,072 weights. So the Qwen result came from a bounded, tool-assisted workflow, not from inspecting raw weights alone.
Results across the three pass budgets
The researchers reported these benchmark scores for Qwen3.5-0.8B, along with the share of MATH-500 answers that reached the model’s 32,768-token limit:
| Qwen3.5-0.8B condition | MATH-500 | GSM8K | IFEval | MATH-500 answers at 32,768-token limit |
| Base model | 7.9% | 13.8% | 21.0% | 93.0% |
| Edited, 10 forward passes | 12.3% | 25.7% | 49.0% | 87.2% |
| Edited, 25 forward passes | 14.1% | 25.1% | 39.9% | 86.4% |
| Edited, 50 forward passes | 48.1% | 67.1% | 62.3% | 33.1% |
The 50-pass condition had the highest reported MATH-500 and GSM8K scores, and the lowest MATH-500 token-limit rate. The intermediate results did not rise steadily across every benchmark: for example, GSM8K and IFEval were both lower at 25 passes than at 10.
The biggest reported gains came with the largest tested pass budget, alongside a sharp drop in answers reaching the output limit.
Why the researchers describe it as a loop repair
The researchers said the edit reduced repetitive output that had prevented Qwen from completing answers. In their account, the score gains mainly reflected more answers reaching a conclusion, rather than new mathematical knowledge. The drop in MATH-500 answers hitting the token limit, from 93.0% in the base model to 33.1% after 50 passes, is part of that explanation.
A separate attempt on Gemma 3 270M did not establish a math improvement. Its MATH-500 scores were 6.8% for the base model, then 7.0%, 7.6%, 7.7% and 7.0% after 5, 10, 25 and 50 passes. Gemma used different sampling settings from Qwen.
What the result covered
The reported edit targeted Qwen3.5-0.8B. The experiment did not include Astra editing its own weights or carrying out a cycle of successive autonomous improvements. The researchers discussed recursive self-improvement as a possible future direction.
