MIT AI Model Directly Edits Another AI's Parameters to Enhance Math Skills

Researchers at MIT's CSAIL demonstrated OpenAI's GPT-6 Astra directly modifying the weights of Alibaba's Qwen3.5-0.8B. This "weight-space programming" significantly improved Qwen3.5-0.8B's mathematical reasoning, boosting its MATH-500 score from 7.9% to 48.1% and GSM8K performance from 13.8% to 67.1%. This method offers a new approach to AI model improvement beyond traditional gradient-based training.