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

Humans can reconfigure how they act on the world from a single verbal instruction. Central to this capacity is the ability to represent what an instruction specifies and to bring that content to bear on incoming sensory evidence. However, little is known about where in the human brain instruction content is encoded, how it combines with sensory evidence to produce a choice, and how this integration relates to behavior. Here, we characterized single-neuron and population responses recorded in the ventral temporal cortex, medial temporal lobe, medial frontal cortex and ventromedial prefrontal cortex of neurosurgical patients performing a verbal instruction task. On each trial, a written instruction specified which stimulus feature was relevant and whether the correct response was to choose or to reject the stimulus that best matched it, so that the same image pair required opposite responses under different instructions. Only medial frontal populations jointly encoded the instructed task, and the target feature, whereas ventral temporal populations encoded stimulus features along axes that generalized across instruction conditions. Choice representation emerged first in ventromedial prefrontal and then in medial frontal cortex during stimulus presentation, before a motor response could be prepared. Shuffling instruction content selectively degraded prediction of medial frontal population activity, whereas shuffling stimulus identity selectively degraded ventral temporal activity. To ask how these representations could be computed, we trained a vision-language model with separable visual and linguistic streams converging on a recurrent decision module to perform the same task. Its visual embeddings predicted held-out ventral temporal activity and its recurrent activations revealed a task solution space that predicted medial frontal activity, and the strength of this alignment tracked behavioral accuracy. These results identify a cortical hierarchy in which medial frontal circuitry represents the content of verbal goals and acts on a ventral temporal substrate whose representational axes remain stable, and indicate that functional hierarchies discovered in artificial networks can serve as predictive models of human neural circuitry.