Human medial frontal neurons represent integrated semantic content and sensory evidence to inform decision-making

Medial frontal cortex in neurosurgical patients encodes instructed task content and target features, and its alignment with a vision-language model predicts behavioral accuracy.

• Why it matters: The study addresses where verbal instruction content is represented, how it integrates with sensory evidence to guide choices, and how these neural computations relate to performance.
• What they did: Single-neuron and population activity were recorded across ventral temporal cortex, medial temporal lobe, medial frontal cortex, and ventromedial prefrontal cortex while patients followed written instructions that flipped choose vs reject for the same image pair. Instruction shuffling selectively degraded medial frontal prediction, while stimulus shuffling selectively degraded ventral temporal activity; a vision-language model with separable streams converging on a recurrent decision module was trained to mirror the task.
• The result: Choice-related signals emerged first in ventromedial prefrontal and then in medial frontal cortex during stimulus presentation, and model alignment with medial frontal activity tracked behavioral accuracy.

AI summary of the abstract · may contain errors · not medical advice · read the paper.