“Forget retraining - AI can now tweak another AI's weights”

Key points

• MIT CSAIL researchers demonstrate frontier LLMs can edit neural network weights directly.
• GPT-6 Astra edited Qwen3.5-0.8B weights without gradient descent training.
• Weight changes were 0.016 percent of weights, boosting math performance up to sixfold.
• Improvements occurred on MATH-500 and IFEval benchmarks without retraining.
• Researchers call the approach Weight-Space Programming and seek benchmarks for evaluation.