MIT authors report GPT-6 Astra’s Qwen weight edit

On Nov. 2, MIT CSAIL authors reported that GPT-6 Astra edited the weights of Qwen3.5-0.8B. In their 50-pass condition, the model’s MATH-500 score rose from 7.9% to 48.1%. The authors attribute the main change to fewer repetitive loops that had cut answers short, not to newly added mathematical knowledge.

What the weight edit changed

The authors called their approach Weight-Space Programming. For the Qwen experiment, GPT-6 Astra used a Python kernel to inspect and modify model weights. The edit subtracted a repetition-associated direction from 64 output columns in the model’s multilayer perceptrons, or MLPs, in each of layers 15 and 19. In total, it changed 131,072 weights, about 0.016% of the model.

This was not a blind edit to untouched weights. Astra also ran Qwen through forward passes, which send inputs through a model to produce outputs. The authors used no gradient information or training loop to calculate the edit, but the target model was run repeatedly to examine and test its behavior.

Why the number of passes mattered

The authors tested budgets of 10, 25, and 50 forward passes. On MATH-500, Qwen3.5-0.8B scored 12.3% after 10 passes and 14.1% after 25, compared with 48.1% after 50. The share of MATH-500 answers reaching the 32,768-token output limit was 87.2% and 86.4% in the first two conditions, then 33.1% after 50 passes; the baseline was 93.0%.

Astra had access to 7,469 GSM8K and 7,275 MATH training-set examples for practice, according to the authors. They say the evaluation sets were not shown to Astra. The larger pass budget and repeated model runs were part of the setup behind the strongest reported result.

The reported Qwen scores

In the 50-pass condition, the authors reported these baseline and edited scores for Qwen3.5-0.8B:

The authors’ breakdown counted 579 rescued outputs and five regressions. Of the rescued answers, 576 had previously been truncated; accuracy on answers that had already finished was unchanged, according to the authors. That pattern supports their explanation that the edit mainly helped Qwen complete answers rather than improving its underlying mathematical ability.

What the Gemma test adds

The authors also tested Gemma 3 270M, which rarely reached its output limit. Its MATH-500 baseline was 6.8%; edited results ranged from 7.0% to 7.7%, and the authors did not establish a mathematical gain. In their comparison, 8.5% of Gemma answers reached the token limit, versus 93% of Qwen’s baseline answers. The contrast fits the authors’ account of the Qwen result: reducing a frequent looping problem can let more answers finish, while the same kind of edit did not produce an established math-score gain in a model that rarely hit its limit.