Inhibitory Cross-Talk Enables Functional Lateralization in Attention-Coupled Latent Memory

Researcher Hong Jeong has published a groundbreaking paper introducing a brain-inspired transformer architecture that achieves functional lateralization through inhibitory cross-talk between memory banks. The core innovation is a memory-augmented transformer where attention serves simultaneously as retrieval, consolidation, and write-back operator through the $A^\top A V W$ update, creating a principled projection from observation space to latent memory to supervised transformation. The architecture partitions memory into lateralized left and right banks coupled through a sign-controlled cross-talk matrix $W_s$, with inhibitory coupling ($s=-1$) mimicking the net inhibitory effect of callosal projections in human cortex.

On a controlled symbolic benchmark combining episodic bijection cipher recall with arithmetic progression rule extraction, the inhibitory model demonstrated remarkable specialization: it reduced cipher-domain loss by 124× over baseline while matching performance on the arithmetic domain. The research confirms that excitatory cross-talk causes bank-dominance collapse where one bank monopolizes inputs, while inhibitory coupling achieves saturated specialization with $\mathcal{D}_{sep} = \pm 1.00$ and $\mathcal{P}_{ct} \approx 0$. This work provides both a mathematical framework for understanding memory lateralization and a practical architecture that could significantly improve AI systems requiring both episodic memory and rule-based reasoning.