Editing 0.016% of a model's weights reduces looping on math tasks

Researchers tested a targeted change to the weights of Qwen 3.5, an 800-million-parameter model that was getting stuck in repetitive loops on math exercises. The edit was made without any new training or backpropagation. On the MATH-500 benchmark, the authors report a score rising from 7.9% to 48.1%, and a later public re-evaluation on Hugging Face reported 51%. The result depends on a specific protocol, and reducing generation loops does not by itself prove a general improvement in the model's mathematical ability. A second experiment is expected to add nuance to these findings.