The neurocognitive variational autoencoder (NCVA) recovers trial-level drift-diffusion parameters from EEG while reconstructing the recorded signal, and has been validated on a rapid perceptual discrimination task. Whether it transfers beyond that setting is unknown. The present study applies it to an abstract reasoning task in which participants complete sequences of icons governed by an implicit rule — a paradigm with reaction times an order of magnitude longer and no frequency tagging to supply the encoder with a known neural marker.

A single model was trained across 19 participants, using a learnable subject embedding to accommodate anatomical differences. The generative pathway transferred: waveforms reconstructed from EEG corresponded closely to the empirical event-related potentials at centro-parietal sites (median r = +0.84). The discriminative pathway did not. Within participants, the drift rate inferred from EEG carried almost no trial-to-trial information about reaction time, and perturbing it did not produce the predicted change in the generated waveform.

Three explanations were tested and none survives. Pooling controlled overfitting without strengthening the within-participant relationship, so the result does not reduce to sample size. The recovered parameters were free to vary by participant, so a population-level constraint cannot account for it. And fitting the diffusion model directly to behaviour showed that it reproduces the observed reaction-time distributions closely when all three parameters are estimated freely — though only by assigning most of the response interval to non-decision time and placing the boundary at the edge of its admissible range.

The generative side works in this setting, but the diffusion model does not carry usable trial-level information, and the reason has not been isolated. Whether the signal is absent in this form, obscured by stimulus-locked analysis, or missed by the architecture cannot be distinguished from these data.