UK — The Register
Forget retraining - AI can now tweak another AI's weights
AI1 min read
MIT researchers say a frontier AI model can edit the numbers inside another AI model directly, instead of retraining it. They call this approach Weight-Space Programming.
A new way to change an AI
Researchers at MIT's CSAIL lab found that a powerful AI model, called GPT-6 Astra, can rewrite the weights of another model. Weights are the numbers that hold what a model has learned. Usually, changing those numbers means retraining the model. In this case, the AI simply read the numbers, understood what they did, and changed them.
Key numbers
- Astra changed only 0.016 percent of the target model's weights.
- Math scores on the MATH-500 test rose by as much as six times.
- Instruction-following scores on the IFEval test rose about three times.
- No training with gradient descent, the usual learning method, was used.
- The target model was Qwen3.5-0.8B.
What are weights?
A neural network is a type of AI built from many connected parts. Each connection has a number called a weight. Together, these weights decide how the model answers. Until now, people mostly changed weights by training the model on lots of data. This study shows that an AI can also read and edit weights like readable code.
Why it matters
If this works well, AI models could be fixed or improved without the usual costly training. The researchers say it may also help avoid catastrophic forgetting, where a model loses old skills when it learns new ones. The method may also help with understanding models and making them safer. These are early signs, and the team plans to build a standard test to check the idea.
What to watch
- The team plans to build a benchmark and a harness for Weight-Space Programming on language models.
- Researchers will check if the method works on more models and tasks.
- Experts will look for proof that it does not cause new problems.
- The idea of recursive self-improvement, where AI improves itself, will get more attention.
Glossary
- weights
- The numbers inside an AI model that store what it has learned and control how it responds.
- GPT-6 Astra
- The OpenAI AI model that the researchers asked to edit another model's weights.
- Qwen3.5-0.8B
- A small AI language model that was edited in the study to improve its math skills.
- catastrophic forgetting
- When an AI model loses skills it had before after learning something new.
- benchmark
- A standard test used to measure and compare how well AI models perform.