On the extended NASCAR synthetic task, where the true number of discrete states shifts from 4 to 6 mid-trajectory, the RS LDS model (with k=6 allocated) still maintains only 4 active states after the dynamics switch, as seen in its most-likely discrete state trajectories. On real IBL mouse electrophysiology data, the RS LDS shows much smaller fluctuations in the number of occupied states per trial (spline r²=0.16) compared to the IRSLDS (r²=0.46), indicating it misses trial-varying structure in state usage. The limitation is the fixed cardinality of the discrete state space in the HMM-based transition prior.
Evidence
correlational
Key metric
Extended NASCAR (k=6, 5 seeds): RS LDS test LL 8015.93, dyn MSE 0.35; IBL per-trial spline r²=0.16 (RS LDS) vs 0.46 (IRSLDS); IBL per-trial k=8 test LL: RS LDS −1399.030
Caveat
The IBL results are from a single session (session id ae8787b1-4229-4d56-b0c2-566b61a25b77) and the authors note they do not aim to make general claims about neural computation.