LLMs and Human Cognition Show Deep Structural Similarities

A recent study published on arXiv disputes the notion that large language models (LLMs) are entirely distinct forms of intelligence. The researchers contend that, although there are variations in physical makeup, learning experiences, and surroundings, modern LLM systems align with human cognitive processes on essential organizational aspects. They highlight structural parallels across five areas: inferential organization, computational design, representational framework, prediction-based learning, and reinforcement-learning-like strategies that facilitate goal-oriented behavior. These results bolster a more expansive understanding of intelligent cognition, indicating that the observed similarities are not just anthropomorphic interpretations but signify true convergence.