MIT Researchers Raise Qwen Math Score to 48.1% With Weight Edits

Researchers affiliated with the Massachusetts Institute of Technology’s Computer Science and Artificial Intelligence Laboratory used OpenAI’s GPT-6 Astra to edit Alibaba’s Qwen3.5-0.8B model, raising its MATH-500 score to 48.1% from 7.9% in an experiment with 50 forward passes.

The method, which the researchers call “Weight-Space Programming,” lets a language model inspect and change another model’s weights directly, without training or gradient updates. The edit touched 131,072 weight values, equal to 0.016% of Qwen’s weights, and targeted repeated reasoning loops that caused answers to run into a 32,768-token limit.

The edit reduced those cutoffs but did not eliminate them. On MATH-500, the share of answers reaching the token limit fell to 33.1% from 93%. In an initial set of 96 practice answers, 85 had hit the limit before the edit.

The score depended on the execution budget. Separate tests with 10 and 25 forward passes produced MATH-500 scores of 12.3% and 14.1%, respectively, compared with 48.1% at 50 passes. The results suggest the edit helped Qwen finish answers that had been cut short, rather than adding a new mathematical skill.

A separate paired evaluation on 1,152 held-out practice problems recorded 579 rescued answers and five regressions. Of the rescued answers, 576 had been truncated before the edit. Accuracy on answers Qwen had already completed was unchanged.

The approach showed little effect on Google’s Gemma 3 270M, which rarely hit the token limit. Its MATH-500 score rose from 6.8% to a maximum of 7.7% at 25 forward passes. Astra reported that the gain had not been established.

Four MIT CSAIL-affiliated researchers posted a blog describing the early results on Nov. 2. The work had not been peer reviewed. Ahsen Khaliq rebuilt the edited Qwen model from the experiment logs and reran the evaluation, but the results have not been independently replicated. The researchers are developing a benchmark and harness for the method.