Next-gen neuro-AI: Aligning foundation models and biologically plausible architectures with human cognition

Cognitive computational neuroscience has entered a transformative era. The rapid rise of large multimodal foundation models, state-space architectures, and advanced transformers has shifted the conversation: we are no longer just building artificial intelligence to solve tasks, but actively using these models as candidate mathematical "theories of the mind." By comparing the latent representations of artificial neural networks with biological neural activity, researchers are uncovering unprecedented insights into human perception, memory, decision-making, and language.However, as AI systems grow more capable, the alignment gap between biological and artificial systems persists. Artificial systems still rely on energy-intensive backpropagation, lack the structural and metabolic constraints of the human brain, and struggle with continuous, lifelong learning without catastrophic forgetting. To bridge this gap, next-generation Neuro-AI must move beyond abstract machine learning to incorporate biological plausibility—such as predictive coding, synaptic plasticity rules, neuromorphic efficiency, and active inference. Furthermore, as both the brain and advanced AI remain complex "black boxes," there is an urgent need for novel mechanistic interpretability methods to map artificial latent spaces directly to human neuroimaging and behavioral data.This Research Topic aims to gather pioneering empirical, theoretical, and methodological work at the intersection of human neuroscience and artificial intelligence. We seek contributions that advance the development, validation, and interpretability of computational models inspired by biological brain mechanisms. Ultimately, this collection aims to foster a bidirectional exchange: using insights from human neuroscience to engineer more efficient, robust, and aligned AI, while using advanced AI architectures to simulate and decode the complexities of human cognition.We welcome submissions (Original Research, Reviews, Methods, and Perspective articles) addressing, but not limited to, the following modern themes:• Foundation Models as Theories of Mind: Utilizing LLMs, LMMs, and state-space models to simulate and explain human cognitive processes.• Biologically Plausible Learning: Implementing learning rules (e.g., predictive coding, Hebbian plasticity, forward-forward algorithms) that bypass traditional backpropagation.• Mechanistic Alignment & Benchmarking: Bridging the representation gap by benchmarking AI models directly against human fMRI, EEG/MEG, and behavioral datasets.• Brain-Inspired Efficiency & Neuromorphic Computing: Computational models exploring how the human brain's low-power, sparse, and continuous learning capabilities can be translated to AI.• Embodied AI and Active Inference: Simulating sensorimotor loops, spatial navigation, and decision-making within virtual or physical environments.• Model Interpretability and Explanability: Developing novel tools to probe and understand the inner workings of deep Neuro-AI architectures in comparison to human neural circuits.• Cognitive "Digital Twins" and Closed-Loop Systems: Using computational models to predict individual cognitive states or guide personalized brain stimulation/rehabilitation. Topic editors: Itamar Lerner, Matteo Toscani Source: https://www.frontiersin.org/research-topics/82982/next-gen-neuro-ai-aligning-foundation-models-and-biologically-plausible-architectures-with-human-cognition