NeuroHorizon: Long-Horizon Forward Prediction of Neural Population Activity via Autoregressive Decoding with Hierarchical Memory

Forward prediction of neural population activity is a prerequisite for closed-loop brain-computer interfaces and a stringent test of learned neural dynamics. Unlike neural decoding or masked reconstruction, this task requires a model to forecast future spiking activity from neural history alone and to remain stable when its own predictions become part of the conditioning context. We introduce NeuroHorizon, an encoder-decoder architecture that combines event-level tokenization of spike trains with an autoregressive causal decoder for population-level firing-rate prediction. The decoder uses a hierarchical tail+segment memory to retain recent predictions at high resolution while compressing longer prediction history, and scheduled sampling reduces the mismatch between teacher-forced training and autoregressive rollout. Evaluated on a multi-horizon motor-cortex benchmark with additional motor- and visual-cortex datasets for scaling and cross-population transfer, NeuroHorizon consistently outperforms strong baselines and remains stable under multi-step rollout where prior approaches collapse. These results show that long-horizon neural forecasting becomes feasible when models are explicitly trained and evaluated under autoregressive rollout, providing a foundation for closed-loop BCI and multi-step neural-state prediction.