MIT CSAIL researchers found frontier LLMs can read and write neural network weights directly. GPT-6 Astra identified the game a network was trained on from its raw weights alone. It also fixed a failing car-control network by editing 3.5% of its weights. Astra improved Qwen3.5-0.8B’s math performance by editing 0.016% of its weights. This raised MATH-500 scores from 7.9% to 48.1% and tripled IFEval results. Astra wrote small game-playing networks from scratch that outperformed PPO-trained ones. The approach works without environment feedback or training loops. Researchers call this Weight-Space Programming. It suggests LLMs may soon natively optimize neural networks through reasoning. The work covers reading, fixing, and writing weights across multiple experiments. All results were achieved using natural language instructions only.
What changed
Neural network weights were optimized only by symbolic algorithms like Adam or gradient descent.
What it unlocks
Directly reading, writing, and editing neural network weights using natural language instructions without training loops or gradient descent.
- 17,253 weights per network
- 3.5% of weights edited
- 0.016% of Qwen3.5-0.8B weights edited
- 7.9% to 48.1% MATH-500 improvement
What you need to act on it
- access to frontier LLMs like GPT-6 Astra
- raw neural network weights provided as text input
Sources