MIT: one AI fixed another AI's math skills without retraining it

MIT researchers report that one AI fixed another AI's math problems by editing it directly, with no retraining. The fixer was OpenAI's GPT-6 Astra. The patient was a small Qwen model from Alibaba, and its math score jumped from 7.9% to 48.1% after the edit.

What changed

An AI model's behavior is stored in weights, the numbers that decide how it acts. Weights are normally set through training. Even small models have far too many to read or adjust by hand. In this research, Astra skipped the training and rewrote some of the weights directly.

The Qwen result

The Qwen model had a habit of repeating itself, which dragged down its math results. Astra changed just 0.016% of the model's weights to cut that habit. That tiny edit was enough to take the score from 7.9% to 48.1%.

It did not work everywhere. The same approach did little for Google's smaller Gemma 3, and its math score barely moved. Gemma 3 lacked one clear flaw like Qwen's repetition, so there was less for Astra to fix.

Other tests

• Astra named the game an AI had learned, using only that AI's weights.
• Astra wrote a simple Breakout player from scratch, without training.

Why people are watching

The authors say this could one day let AI upgrade itself, add or remove skills in other models on request, or check them for hidden problems. The Rundown AI notes that top labs are racing toward automated AI researchers. This is one small example, but if models can improve each other, AI progress could speed up. That could mean better tools sooner, and the risks of that speed deserve attention too.

What is not known

The material does not say whether this works on larger models or when anything like it could reach everyday products. Only some models improved, and Gemma 3 barely did. For now, this is early research, not a tool anyone can use.