Abstract
Monolithic world models predict the entire next state at every step, spending capacity re-predicting the static majority of a scene and injecting error into it. We ask whether explicitly modeling what changes — a per-object change gate plus a residual delta head that perturbs only gated objects, copying the rest verbatim — is a more effective and interpretable bias.
On a MuJoCo tabletop pushing benchmark (3–8 objects), the sparse/residual model is 2.5–4.6× more accurate than a dense MLP at 8.6–11.1× fewer parameters overall, but this headline is driven entirely by error suppression on the static majority: at N=5, overall L2 is 0.104 (sparse) versus 0.107 (no-op), and changed-object L2 is 0.426 versus 0.446 — at or barely above the floor of predicting no motion. The durable win is therefore change detection that avoids error injection (F1 0.80–0.87, precision 0.92–0.97) with delta regression at roughly no-op quality, not accurate dynamics on movers. The gate transfers across object counts (99.4% F1 retention) and is more sample-efficient. For planning, once featurized for planner-visited states, sparse reaches 0.23 ± 0.06 success versus random 0.15 and dense 0.00 at every seed.