LLMs Natively Read and Write Neural Network Weights

Researchers introduced 'Weight-Space Programming,' demonstrating that frontier language models can natively interpret and edit raw neural network weights without traditional gradient training loops. Using GPT-6 Astra, models successfully wrote game-playing networks from scratch and fixed reasoning flaws in Qwen3.5-0.8B by directly editing a tiny fraction of its weights. This breakthrough represents a paradigm shift from simplistic symbolic optimizers like Adam to AI-driven weight optimization, potentially transforming mechanistic interpretability and reducing model optimization costs. It opens up new possibilities for zero-data capability steering, model auditing without evaluation, and direct recursive self-improvement. By editing just 0.016% of Qwen3.5-0.8B's weights, Astra raised its MATH-500 score from 7.9% to 48.1% and tripled IFEval instruction-following performance under identical sampling settings. This approach allows models to reason about what each part of a neural network computes and adjust weights accordingly to avoid catastrophic forgetting.