Role-agnostic latent decoding of ACC and caudate multi-unit activity dissociates target-feature information across learning

When several features of an object could be the one that is rewarded, learning requires prioritizing the relevant feature over distracting ones. We asked whether the anterior cingulate cortex (ACC) and the head of the caudate carry information about the rewarded (target) feature while its relevance is still being learned. Using a gated-recurrent variational autoencoder whose classifier heads report the features of the chosen object, we decoded multi-unit activity from \Ncd{} sessions in two macaques performing a feature-reward learning task. The roles of features (target versus distractor), the learning stage, and trial outcomes never enter the training objective or labels; attentional filtering is instead read out post hoc by partitioning held-out predictions by feature role. Area contributions are quantified as exact area-level Shapley shares of each decoder's chance-corrected accuracy, and contrasts are tested with session-level permutation tests. After learning, the caudate's target-feature share was above chance throughout the analysis window and exceeded the share for the same trials' distractor features. The ACC's peak share was larger during learning than after, and among the trial bins around the learning criterion only the earliest was above chance; after learning, the ACC's peak share was larger for the features of erroneously chosen objects than for the target. The attribution-and-inference recipe needs only per-area occlusion accuracies and session identifiers, so it should apply unchanged to any multi-area decoder, pretrained or not.