Reasoning with Brain-Like Specialization

Abstract

Human cognition relies on specialized brain networks for language, logic, and social reasoning. Inspired by this organization, we introduce Mixture of Cognitive Reasoners (MiCRo) — a modular transformer that develops brain-like specialization across experts. Each expert aligns with a distinct cognitive domain, enabling interpretable behavior and controllable reasoning. These experts are causally meaningful: ablating one selectively impairs its domain, while routing tokens toward specific experts steers the model’s outputs. MiCRo matches or surpasses standard baselines on reasoning and human behavioral benchmarks, demonstrating that cognitive modularity can enhance both performance and interpretability.

Speaker Bio

Badr AlKhamissi is a PhD candidate at EPFL, working with Antoine Bosselut and Martin Schrimpf. His research bridges machine learning, neuroscience, and cognitive science, focusing on building models aligned with the human brain and behavior. Before joining EPFL, he was an AI Resident at Meta AI and a Research Intern at Sony AI. He holds an MSc in Computational Cognitive Neuroscience from Goldsmiths, University of London, and a BSc in Computer Science from the American University in Cairo. Badr is also active in the research community, serving on program and organizing committees for several conferences and workshops.

Recording