[ITmedia Business Online] New method of "AI directly fixing AI" demonstrated by Massachusetts Institute of Technology with 6 times the correct answer rate without relearning

An AI examines the inside of another AI, finds defects, and fixes them. A research team at the Massachusetts Institute of Technology (MIT) announced such a new AI improvement method on November 2nd.

When they used OpenAI's GPT-6 Astra to rewrite just 0.016% of the numbers inside China's Alibaba small AI model Qwen3.5-0.8B, the correct answer rate for math problems jumped from 7.9% to 48.1%.

What is noteworthy is that the performance was improved without retraining the AI ​​model. Another AI analyzes the internals of the AI, which previously made adjustments using large amounts of data, and directly corrects problematic areas.

The research team called this method "Weight-Space Programming." This research shows new possibilities for AI to improve AI.

A new method to rewrite “800 million numbers”

Inside an AI model, there are huge numbers called "weights." In the human brain, this corresponds to the strength of the connections between nerve cells. Through learning, AI adjusts these numbers and acquires the ability to understand words and solve problems.

Until now, the common way to improve AI performance was to provide additional training data and readjust the weights through extensive calculations. Human engineers may change the design or learning method of a model, but it is not realistic to decipher and modify hundreds of millions or trillions of weights one by one.

A research team at MIT challenged this conventional wisdom. Researchers Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab from the Computer Science and Artificial Intelligence Laboratory (CSAIL) had GPT-6 Astra examine the weights of different AI models and figure out what changes could improve performance.

Qwen3.5-0.8B, which was the subject of the experiment, is a small language model with approximately 800 million weights. Astra discovered that when this model solved mathematical problems, it repeated the same reasoning over and over again, leading to problems it could not come up with an answer to.

So Astra modified certain weights in layers 15 and 19 of the model. The number of changes was 131,072, which is only 0.016% of the total.

As a result, the correct answer rate for the mathematics evaluation test ``MATH-500'' rose from 7.9% to 48.1%. Another math test, GSM8K, improved from 13.8% to 67.1%. Furthermore, the score on the IFEval, which measures the ability to follow instructions correctly, increased from 21.0% to 62.3%.

It's not like I'm inculcating new mathematical knowledge. By removing the cause of AI's inability to arrive at an answer, we were able to unleash its original abilities.