Thinking Language Models with Hidden Cases Can Give Better Result

Big language models (LLM) normally think by typing step words. However, the new research published in arXiv reveals that models can be considered secret in internal mathematical layers, not by producing text. This method can also be transferred to information between agents.

The main findings of the work done can be summarized as follows:

• Metin’s output problem: The current educational method models optimize only for the accuracy of the final answer, do not control the thought process.

-Dört critical failure type: The research proved mathematically and experimental that only leads to four categories of errors such as the collapse of text-based education.

• The right probability of answer falls: These failures reduce the possibility of reaching the correct conclusion of the model.

-**: The models can get more reliable results if it allows you to get rid of text restriction and free calculation in internal layers.

The research refers to the “fall” format of artificial intelligence systems. As people think of speech, models can also treat more rational with internal calculations without output. These findings may be an important guide in the design of the next generation LLM architectures.