Researchers from Peking University and Tsinghua University found some similarities between how the human brain and large language models process logical tasks. Then they went further and used data on brain activity during thinking to improve AI responses.
Researchers from Peking University and Tsinghua University found some similarities between how the human brain and large language models process logical tasks. Then they went further and used data on brain activity during thinking to improve AI responses.
The research was published in the journal Nature Machine Intelligence.
This is deductive thinking — when you need to draw a logical conclusion from several known facts. For example: if all cats are animals, and Murchik is a cat, then Murchik is an animal. For the experiment, the researchers used an open fMRI dataset — brain scans of people who solved such logical problems. Then, they gave the same tasks to 10 language models with the number of parameters from 1.5 billion to 72 billion.
Scientists compared what happens in the parts of the human brain associated with logical thinking with the internal activity of neural networks. It turned out that there are some similarities between the two, although the models and people process information in completely different ways.
The researchers then created a method that uses human brain activity patterns as a guide and slightly changes the inner workings of the LLM as it solves problems. It gives the model an extra hint, not about which direction to process information, but rather about what answer to give.
The method was tested on all 10 models. According to the authors of the work, it improved results in logic tasks and was even able to work on slightly different types of tasks than those for which it was tuned. In some experiments, accuracy increased by up to 13 percentage points.
The researchers also want to test this approach on other cognitive abilities and with other ways of recording brain activity. They suggest that such data could be particularly useful in situations where it is difficult to train a model on text examples alone.
Previously, dev.ua wrote about the UCLA system, which allows you to control a robotic hand or cursor using brain signals.
