Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab report that GPT-6 Astra edited 131,072 weights in Qwen3.5-0.8B and raised its MATH-500 score from 7.9% to 48.1% after 50 forward passes. Their analysis points to a specific explanation: 576 of 579 rescued answers had previously been cut off, while accuracy on answers that had already finished stayed the same. The result is consistent with fewer incomplete answers, not proof that the edit gave Qwen new math knowledge.
What changed in Qwen3.5-0.8B
The authors described the method as “Weight-Space Programming”: using a language model to inspect and directly edit a neural network’s weights, rather than making the changes through a conventional training loop. In this experiment, GPT-6 Astra used a Python kernel to work with the weights, without a GPU or gradient information.
The target was Qwen3.5-0.8B in thinking mode, a 24-layer model with 0.8 billion weights. In the 50-pass run, Astra changed 64 output columns in the model’s multilayer perceptron (MLP) in each of layers 15 and 19. The authors say that amounted to 131,072 weights, or 0.016% of the model. GPT-6 Astra carried out the edit; NeoTeo has covered the model separately.
The reported benchmark results
The authors report gains on all three benchmarks in their comparison of Qwen3.5-0.8B before editing and after 50 forward passes. A forward pass is one run of the model to produce an output.
| Benchmark | Before editing | After 50 forward passes |
| MATH-500 | 7.9% | 48.1% |
| GSM8K | 13.8% | 67.1% |
| IFEval | 21.0% | 62.3% |
The displayed MATH-500 scores differ by 40.2 percentage points. The size of the jump alone, however, does not show whether the model learned new mathematical reasoning; the authors’ analysis of the answers helps explain what changed.
Why the MATH-500 score rose
The authors report that 576 of the 579 outputs their edit rescued had originally been truncated at the 32,768-token limit. They also say aggregate accuracy on answers that had already finished did not change. That pattern points to fewer answers getting trapped in repetitive reasoning and cut off before completion, rather than a demonstrated increase in the model’s underlying math knowledge.
The token-limit results follow the same pattern. On MATH-500, the share of Qwen answers reaching the limit fell from 93.0% before editing to 87.2% after 10 passes, 86.4% after 25 passes and 33.1% after 50. The corresponding MATH-500 scores were 12.3%, 14.1% and 48.1% after 10, 25 and 50 passes, respectively. The sharpest reported change came with the largest editing budget.
Test conditions and the Gemma control
Astra could use 7,469 GSM8K and 7,275 MATH training-set problems as practice material. The authors say it did not see the final evaluation sets. Their Qwen evaluations used budgets of 10, 25 or 50 forward passes, and the reported benchmark scores average different numbers of samples per task. The authors also identify context length and inefficiencies in the evaluation harness as practical limits.
The result did not carry over as an established math gain in the authors’ Gemma 3 270M control. Only 8.5% of Gemma’s answers reached the token limit, compared with 93% of Qwen’s baseline answers. Gemma scored 6.8% on MATH-500 at baseline and reached a reported maximum of 7.7% after editing; the authors did not consider that an established improvement.
