
6.1×Math accuracy vs unedited model · Reported by the MIT CSAIL authors
Astra edits Qwen’s weights to reduce looping without retraining
Weight-Space Programming experiments at MIT CSAIL raised Qwen’s MATH-500 math-test accuracy from 7.9% to 48.1%, the researchers report. They published a research blog and released experiment logs, final networks and exact weight edits on GitHub.
Sources
CodeRikiyaT/lm_understand-logs: Complete logs of the experiments in the blog post 'LLMs can understand and write neural network weights'Official releaseSparks of Weight-Space Programming: LLMs can natively read & write neural network weightsWritten by AI from the sources