Shanghai AI Lab 'Weight-Space Programming' Paper Edits 1,023 Params to Lift Qwen3.5 29pp on MATH-500

A new Shanghai AI Lab/Tongji/Fudan paper systematically studies Weight-Space Programming — letting a reasoning agent inspect and directly edit a model's parameters instead of gradient-training it. Across 16 editing trajectories on Qwen3.5-0.8B and 9B, they report accuracy gains of up to 29.45pp on MATH-500, 36.72pp on IFEval and 17.11pp on SciBench, with a single 1,023-parameter edit producing the 29.45pp lift. Cross-task generalization is uneven and six of 16 trajectories reverted to original weights when candidate edits could not be validated.