Can AI fix another AI's brain? MIT researchers use GPT-6 Astra to boost Qwen's maths skills

Can one artificial intelligence fix another AI's brain? A new experiment by researchers at the Massachusetts Institute of Technology (MIT) suggests that it may be possible. In the experiment, they used OpenAI's GPT-6 Astra to identify and fix a problem inside Alibaba's Qwen3.5-0.8B AI model, significantly improving its mathematical performance. And it did all this without having to train Qwen again. According to researchers, Astra simply changed a very small number of the settings inside the AI model to fix the problem and make it smarter. The experiment, detailed in a research blog post titled, “LLMs can natively read & write neural network weights,” was conducted by researchers Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab from MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL). They call the technique 'Weight-Space Programming'. We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! Blog: https://t.co/bOWCcFhCVq— Rikiya Takehi (@rikiyatakehi) November 2, 2026According to researchers, the idea behind the experiment was to see if one advanced AI model could understand how another AI works and directly change its internal settings, known as weights, to fix problems or improve its performance.So what exactly did the researchers do? For the experiment, the researchers gave GPT-6 Astra access to Alibaba's Qwen3.5-0.8B model through a Python environment. They also allowed Astra to test Qwen capabilities and performance using practice maths questions, with a limit of 50 attempts at running the model. However, the researchers did not allow Astra to retrain Qwen or use the usual methods to teach it new things. Instead, Astra had to identify the problem and fix it by directly changing a small part of the model. According to the paper, during the tests, Astra found that Qwen models frequently got stuck repeating the same lines of reasoning instead of solving a maths problem and giving an answer. Sometimes, it kept repeating itself until it reached its maximum response limit of 32,768 tokens.Astra then identified the parts of Qwen's internal system responsible for this behaviour and changed them to stop the repetition. It modified 131,072 weights to fix the problem. Notably, while this number might sound like a lot, it represents just 0.016 per cent of Qwen's total 800 million weights. In other words, Astra changed only a tiny part of the model to fix something that was affecting its answers.Did Astra fixed Qwen? Yes. According to the paper, after Astra made these changes, Qwen's score on the MATH-500 benchmark, which tests mathematical problem-solving, jumped from 7.9 per cent to 48.1 per cent. Its score on another maths test, GSM8K, also rose from 13.8 per cent to 67.1 per cent. According to the researchers, the changes also improved Qwen's ability to follow instructions, with its IFEval score increasing from 21 per cent to 62.3 per cent.However, the researchers clarify that Astra did not teach Qwen any new mathematical concepts. Instead, it fixed a problem that was stopping the model from completing its reasoning and using the abilities it already had.The researchers nevertheless say this is still an early experiment, and the technique has its limitations. In another test, Astra could not identify which of two AI models had been trained to agree with users simply by looking at their weights. It could do so only after the researchers gave it access to information about where the models were focusing their attention.The team has released the logs of its experiments and is now working on a benchmark to test this technique further. While the findings show that one AI can identify and fix problems inside another AI model, it remains to be seen whether the approach will work reliably with much larger and more complex models.- Ends