Q1 Math score jumps from 7.9% to 48.1%… GPT-6 Astra adjusts weights

GPT-6 Astra directly edited 131,072 weights—0.016% of the total—in Qwen 3.5-0.8B, lifting the model's MATH-500 score from 7.9% to 48.1%, according to early experimental results. MIT CSAIL researchers, in a November 2 post, dubbed the approach "Weight-Space Programming," in which one large language model inspects another model's weights and directly modifies select values. The same intervention raised GSM8K performance from 13.8% to 67.1% and instruction-following (IFEval) scores from 21.0% to 62.3%, with the MATH-500 improvement representing roughly a 6.1x gain. The results, based on up to 50 model iterations on Qwen 3.5-0.8B, suggest weight-level editing by AI systems could emerge as a new model-improvement paradigm alongside training and fine-tuning. Watch for replication on larger models and whether MIT CSAIL publishes full methodology and benchmarks.