Latent Music World Model

Long-horizon controllable symbolic music generation requires persistent musical state, explicit actions, and structural organization across recursive rollout. We formulate this problem as plan-conditioned latent world modeling and develop a hierarchical, action-conditioned Music World Model (MWM) with fast segment states and slow phrase-level plans. Recursive MWM deployment creates a predicted-plan structural drift pathway: later planner inputs contain representations produced by earlier model rollouts. Under matched conditions, we compare a predicted-plan (PP) path with a teacher-plan (TT) path. PP recursively uses its predicted executor-facing slow trajectory, whereas TT replaces only that future trajectory with a target-derived reference. This PP–TT diagnostic isolates the additional divergence associated with recursive structural self-conditioning while holding the executor, actions, codec, decoder, and rollout horizon fixed. Across chained long-form rollouts, the discrepancy concentrates in late phrase transitions and is inherited by subsequent planner contexts. We then introduce Boundary-Aware History Correction (BHC), a training-free, parameter-free inference procedure that combines a local context-regularized anchoring correction with a persistent forward-decaying boundary displacement. The intervention leaves the generator frozen and uses no future teacher information at inference time. Across 16–64-bar validation rollouts, song-level decoded and internal trajectory analyses show that BHC suppresses part of the PP–TT divergence, with the clearest gains after structural mismatch has accumulated. Controlled mismatch experiments further show that the recoverable gain increases with the injected gap. These results support BHC as a targeted suppressor of predicted-plan structural drift, rather than a guarantee of global stability or drift elimination.