Inference Structure in Numeric Vectors, Control Clues Also Revealed

According to AI TIMES, Yale University researchers analyzed in a study released on October 19 that symbolic structures resembling human language and logic are formed within the internal vectors of seven large language models. They found that even when a model's internal representations are reconstructed with a mathematical framework that separates role and content, its behavior is largely preserved, and that changing the links between certain concepts can also alter the direction of responses. This suggests that chatbots may appear to process information by simply stringing together numbers, but they may internally distinguish and handle relationships such as sentence components and mathematical positions. In particular, similar patterns were observed not only in language but also in arithmetic, logic, and coding, offering clues to the common scaffolding of reasoning. However, the researchers noted that such structures do not explain all of a model's operation, nor can they be directly expected to lead to consistent control in real-world settings. The study's focus was less on the claim that AI thinks like humans than on making internal representations more interpretable and testing the possibility of intervention.