Forget retraining - AI can now tweak another AI's weights
'Weight-Space Programming' fiddles with nerd knobs to boost benchmark scores
Recent research by MIT's CSAIL has shown that frontier AI models like OpenAI's GPT-6 Astra can directly edit the weights of another model, rather than just indirectly improving through training. This process, dubbed "Weight-Space Programming," allows Astra to change as few as 0.016 percent of Qwen3.5-0.8B's weights, significantly improving its math abilities.
The researchers found that Astra could program the neural network weights of a policy network from scratch, capable of playing Breakout as well as possible. This technique could improve model optimization, reduce catastrophic forgetting, and aid in model interpretability and alignment. The next step involves creating a benchmark for Weight-Space Programming to standardize and accelerate progress in this area.
Written by urgent.news from The Register Science's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
