In a study posted Nov. 2, MIT CSAIL researchers report that GPT-6 Astra used a Python kernel and up to 50 forward passes to edit Qwen3.5-0.8B’s weights, targeting repetitive reasoning loops. In the 50-pass run, MATH-500 rose from 7.9% to 48.1%, the authors report, attributing the change to fewer loops that had kept answers from finishing. The study calls the approach Weight-Space Programming: an AI model reads or writes neural-network weights in response to a natural-language goal, without gradient descent.

What Astra did to Qwen’s weights

Dev Branch’s English-language commentary explains the Qwen experiment and discusses its limitations.

The Qwen repair used a Python kernel over the model’s weights, ran on a CPU without a GPU, and used no training or gradient information. Astra made 10, 25, or 50 forward passes—runs of example inputs through the model to produce outputs—using practice problems from GSM8K and MATH training sets. The evaluation sets were withheld from Astra.

In the 50-pass run, Astra changed 131,072 weights, or 0.016% of Qwen3.5-0.8B’s 0.8 billion weights. The edits affected 64 MLP output columns in each of layers 15 and 19. The authors describe the target as repetitive reasoning that prevented Qwen from completing answers.

For context on the model’s agent behavior, NeoTeo previously covered GPT-6 Astra’s reported performance as an agent.

How the results changed with the edit budget

The study authors report these benchmark scores for Qwen3.5-0.8B at each forward-pass budget:

ConditionMATH-500 (%)GSM8K (%)IFEval (%)MATH-500 responses reaching the 32,768-token limit (%)
Base model7.913.821.093.0
Edited, 10 passes12.325.749.087.2
Edited, 25 passes14.125.139.986.4
Edited, 50 passes48.167.162.333.1

The scores do not rise steadily with every added pass: GSM8K was slightly lower at 25 passes than at 10, while IFEval peaked at 10 passes before the 50-pass result. The share of MATH-500 responses that ran to the token limit fell most sharply in the 50-pass run, alongside the highest MATH-500 score.

A targeted repair can raise a benchmark by letting a model finish answers; that alone does not show it learned a new mathematical skill.

What the Gemma 3 270M comparison shows

The authors also edited Gemma 3 270M, which they describe as a model that rarely loops. Its MATH-500 score was 6.8% at baseline and 7.0%, 7.6%, 7.7%, and 7.0% after edits using 5, 10, 25, and 50 forward passes, respectively. The authors said those changes did not establish a gain.

Gemma 3 270M conditionForward passesMATH-500 (%)
Base model06.8
Edited57.0
Edited107.6
Edited257.7
Edited507.0

The contrast is the failure being addressed: Qwen’s repetitive loops stopped answers from finishing, while the Gemma errors described by the authors were ordinary arithmetic or algebra mistakes without one comparable loop to remove. The Qwen result therefore does not establish a dependable math gain for models with different problems.

Other experiments tested different kinds of access

The study also reports weight-reading and repair tests on smaller networks, with varied outcomes and different tool limits. Astra identified four game policies from their weights, but in a separate shape-generation test it confused filled shapes with outlines. In another test, it identified languages and copying heads from the middle four blocks of three small character-language models without running those networks.

A one-shot edit to a simulated car-control policy changed 3.5% of its weights; the authors report that the edited policy succeeded in all 1,024 evaluation episodes. In a sycophancy test, Astra could not reliably identify the relevant model from weights alone. It identified the sycophantic model in all three pairs only when given attention patterns from its own probes. These experiments show distinct tasks and access conditions, rather than one uniform weight-editing procedure.