Current AI systems struggle with true autonomous learning, failing to adapt like biological organisms, according to a new paper. Researchers propose a hybrid architecture integrating observational learning (System A) and active behavior learning (System B), guided by meta-control signals (System M). This approach aims to overcome AI's static learning limitations by dynamically switching between learning modes based on internal cues, potentially enabling more flexible adaptation to real-world, dynamic environments across timescales.

The framework draws explicit inspiration from cognitive science and animal behavior, addressing a fundamental limitation in contemporary AI development.