BCIwiki (bciwiki.com) — A framework called Task-Conditioned Latent Alignment (TCLA) achieved a mean R² of 0.476±0.014 with a negative-R² failure rate of 6.8% in long-term cross-session decoding across seven nonhuman primate datasets, according to a preprint posted to arXiv on September 23, 2026. The study, by Canyang Zhao, Bolin Peng, J. Patrick Mayo, Ce Ju and Bing Liu, has not been peer reviewed.
Stable long-term decoding remains hard in invasive brain-computer interfaces (BCIs) because the neural populations recorded by implanted electrodes change from session to session. The authors argue that existing latent alignment methods may overlook task-dependent structure when adapting across sessions. TCLA learns a low-dimensional source representation using neural reconstruction and continuous behavioral supervision. When adapting to a new session, it keeps that shared representation fixed and maps target neural activity into the source latent space by aligning source and target distributions separately for each task condition.
Across 1,356 within-subject session pairs, TCLA reached a mean R² of 0.371±0.009 with a 6.8% failure rate. Across 2,134 cross-subject session pairs, it reached 0.218±0.004 with a 12.9% failure rate, which the authors describe as substantially better than comparison methods. They attribute the gains to preserving behaviorally relevant, task-dependent latent structure. The source code is publicly available.