Weight-Space Programming allows direct modification of neural network parameters instead of using gradients. Testing on Qwen3.5-0.8B and 9B shows that modifying just 1,023 parameters can improve MATH-500 accuracy by up to 29.45 percentage points.
HOW THIS AFFECTS YOU
●
builderYou can improve model performance on specific tasks by editing a very small number of parameters directly.
●
researcherYou can explore how localized weight interventions affect cross-task generalization in models like Qwen3.5.