Neural decoding stops needing only paired recordings: BeeMO gains cross-task transfer from 6.5 hours of unpaired behavior

The training data for neural decoding can be split into two streams: scarce paired neural-behavioral recordings and abundant unpaired behavioral trajectories, and the latter can take part directly in pretraining and improve decoding. Strong decoders such as POYO and POSSM previously learned neural representations and behavioral dynamics jointly from paired neural-behavioral recordings, so scaling meant collecting more synchronized recordings. On center-out reaching, random-target reaching and isometric wrist tasks in nonhuman primates, BeeMO-POYO and BeeMO-NDT* with 6.5 hours of additional behavior during pretraining beat their respective multi-session backbones on all tasks; the paired comparisons were significant for BeeMO-POYO on random-target reaching and isometric wrist and for BeeMO-NDT* on random-target reaching, while the isometric-wrist difference for BeeMO-NDT* was smaller and not significant.

The behavior decoder first learns the temporal structure of movement trajectories on its own, and the neural encoder compresses spike activity into latent tokens that are injected into the decoder through cross-attention when needed. The behavior decoder is trained with a behavior-forecasting objective, predicting behavior at target times from past behavioral context; the cross-attention step is skipped when no neural data is present, so one training run can mix behavior-only and paired sessions. During training, behavioral context is masked out for paired sessions with probability 0.5 so that at test time the model must decode from neural activity alone; removing this masking degraded BeeMO-POYO and BeeMO-NDT* by roughly 6 to 9 percent across tasks, falling below even their respective backbone models.

The experiments cover three motor tasks, center-out reaching, random-target reaching and isometric wrist, from intracortical recordings in nonhuman primates, with evaluation on animals not seen during pretraining and with POYO and NDT* as the two neural encoder backbones. The authors propose testing next whether behavioral pretraining still helps on benchmarks spanning brain-wide and multi-region recordings.

The authors note that BeeMO requires an observable temporal target or a meaningful behavioral proxy and therefore does not directly apply to cognitive states without such measurements; the datasets are largely structured, task-driven recordings, and extending to freely moving behavior may require richer behavioral representations. They also report that BeeMO-POYO's gain over POYO stays roughly constant as neural sessions increase, whereas BeeMO-NDT*'s gain widens with more sessions, and the reason for the difference is left for further study.