NeuroCogMap Reveals LLM Cognitive Organization, Links to Human Brain

A team of researchers led by Zhongxiang Sun introduced NeuroCogMap, a cognitive neuroscience-inspired framework that systematically organizes the internal representations of large language models (LLMs) into functional parcels—analogous to brain regions. These parcels form a stable, semantically coherent organization that is partly conserved across different LLM architectures and is functionally linked to model outputs. The framework goes beyond surface-level behavior by mapping how distinct cognitive functions like reasoning, memory, and language processing are distributed within the model’s latent space. Using NeuroCogMap, the team identified reproducible internal signatures corresponding to major LLM failure modes: hallucination, bias, refusal failure, and sycophancy. Each failure type corresponds to a distinct disruption in specific representational or behavioral-control systems, enabling mechanism-guided detection and targeted intervention.

Beyond analyzing LLMs themselves, NeuroCogMap bridges artificial and biological cognition. The framework improves prediction of human cortical responses during naturalistic language comprehension, with the strongest correspondence found in higher-order association cortex—the brain's integrative reasoning hubs. At the cognitive level, NeuroCogMap’s internal signatures expose latent decision-making strategies that refine classical models of human choice. This 79-page study, with 6 main figures and 5 extended figures, establishes a system-level approach for mapping functional organization in artificial systems and relating it directly to human cortical function and cognitive behavior. The work opens new avenues for diagnosing LLM failures and aligning AI systems more closely with human cognition.