The Reflexive Integrated Information Unit: A Differentiable Primitive for Artificial Consciousness
arXiv Preprint Archive June 15, 2025 Gnankan Landry Regis N'guessan, Issa Karambal
A new type of artificial neural network cell, the Reflexive Integrated Information Unit (RIIU), is introduced to make consciousness-like computation trainable and scalable. Each RIIU tracks its own causal influence on the network and learns to maximize local information integration. In a simple grid-world test, a network of RIIUs recovered over 90% of reward within 13 steps after a simulated actuator failure, twice as fast as a comparable gated recurrent unit (GRU). The authors argue this approach transforms a philosophical debate into an empirical mathematical problem.