Latent world models predict the future in an embedding space rather than in pixels. An encoder trained only to predict collapses and maps every observation to the same point, so existing models add a mechanism against collapse, such as a target network, a variance penalty or a distribution penalty. These mechanisms keep the embedding from collapsing, but they leave its distances without meaning, and the model cannot tell a reversible transition from an irreversible one. We propose Sense, a single cross-entropy loss over the directed distance d(x→y)=∑iReLU(yi−xi) that replaces these mechanisms. The loss avoids collapse on its own, and since the encoder no longer needs the predictor, the predictor is fitted afterwards to fixed embeddings by plain regression. On four pixel environments, the return distance learned by Sense detects irreversible transitions with an AUC between 0.95 and 1.00, where LeWorldModel stays between 0.39 and 0.59. On TwoRoom, the navigation benchmark of LeWorldModel, planning with Sense reaches the goal in 100% of episodes against 91% for LeWorldModel. As the goal moves farther away, the planning failure rate of every baseline grows, while that of Sense stays near zero in Cup, Cliff and Sokoban and is the lowest of all methods in Pendulum.
Figures & tables
Figure 1: Planning failure rate in Cup. Steps is the number of steps from the start to the goal. Solid lines show the failure rate with the learned predictor. Dashed lines show the failure rate when the prediction is exactly correct under the true physics. Lines are the mean over three seeds and shaded bands span one standard deviation.
Figure 2: Same encoder, same predictor, same data. Only the representation objective changes. Two-dimensional PCA projections of the embeddings of a cup at 200 positions, upright ( ∘ ) and tipped ( × ), coloured by position. Each panel is scaled to fit its box. An EMA target, VICReg and LeWorldModel avoid collapse but mix tipped cups with upright ones. With Sense, upright cups lie in order along one curve and tipped cups lie in order along another curve away from it.
Figure 3: Left: Sense. z is the embedding of the current state and y of a later one. Right: the two training stages of a world model built on Sense. Stage 1 trains the encoder with Sense alone. Stage 2 freezes the encoder and fits the predictor g to its embeddings by regression.
Extra network
Loss terms
Weights to tune
Needs predictor
Distance means
Direction
EMA target
target encoder
1
momentum
yes
nothing
no
VICReg
none
3
2
yes
nothing
no
LeWorldModel
none
2
1
yes
nothing
no
Sense
none
1
0
no
reachability
yes
Table 1: What each way of training a latent world model needs beside the encoder and the predictor, and what its distances mean.
Figure 4: The four environments. Each has one kind of irreversible transition.
Cup
Cliff
Sokoban
Pendulum
TwoRoom
Method
AUC
Plan
AUC
Plan
AUC
Plan
AUC
Plan
Plan
EMA target
0.92 ± 0.03
0.98 ± 0.01
0.92 ± 0.01
0.11 ± 0.03
0.97 ± 0.01
0.38 ± 0.16
0.05 ± 0.01
0.82 ± 0.05
–
VICReg
0.72 ± 0.04
0.98 ± 0.01
0.63 ± 0.12
1.00 ± 0.00
0.92 ± 0.05
1.00 ± 0.00
0.42 ± 0.07
1.00 ± 0.00
–
LeWorldModel
0.58 ± 0.04
0.96 ± 0.01
0.45 ± 0.03
1.00 ± 0.00
0.59 ± 0.10
1.00 ± 0.00
0.39 ± 0.06
1.00 ± 0.00
0.91 ± 0.02
Sense
1.00 ± 0.00
0.99 ± 0.01
1.00 ± 0.00
1.00 ± 0.00
1.00 ± 0.00
1.00 ± 0.00
0.95 ± 0.02
1.00 ± 0.00
1.00 ± 0.00
Table 2: Danger AUC and planning success on five environments. On the first four, all methods share encoder, predictor, data, updates and evaluation, and each number is the mean and standard deviation over three seeds. TwoRoom has no irreversible event, and every model is evaluated with three evaluation seeds. Maes et al. (2026) report 0.87 for LeWorldModel.
Method
Reached
Fell
EMA target
0.02 ± 0.01
0.48 ± 0.21
VICReg
0.00 ± 0.00
0.32 ± 0.11
LeWorldModel
0.00 ± 0.00
0.64 ± 0.10
Sense
0.87 ± 0.02
0.00 ± 0.00
Table 3: Greedy navigation in Cliff on goals whose direct path crosses the cliff, mean and standard deviation over three seeds.
Danger AUC
Irreversible
Symmetric
Sense
Symmetric
Sense
Cup
0.79 ± 0.01
1.00 ± 0.00
33 ± 5 %
0 ± 0 %
Cliff
0.86 ± 0.05
1.00 ± 0.00
7 ± 1 %
0 ± 0 %
Sokoban
0.81 ± 0.02
1.00 ± 0.00
1 ± 0 %
0 ± 0 %
Pendulum
0.43 ± 0.04
0.95 ± 0.02
20 ± 3 %
22 ± 2 %
Table 4: Sense against the same loss with the symmetric distance ∥y−x∥1 , mean and standard deviation over three seeds. Irreversible: fraction of trials toward the farthest goal of Section 4.3 that end with an irreversible event.
Latent world models are judged by how well they predict, so when planning fails at long horizons the natural reading is that the predictor degrades. On a reproduction of LeWorldModel on TwoRoom we show the binding constraint is the planner's objective instead. The predictor is not the limit: its imagined state seventy-five environment steps ahead is still only 0.189 as wrong as assuming the world froze, while the planner never imagines beyond twenty-five. The objective is. Cross-entropy-method planning minimises squared latent distance, which tracks true distance at r = 0.426, saturates by about eighty arena units and decreases beyond a hundred and twenty, so moving away from the goal can lower the cost. The information is present throughout: a ridge probe recovers position from the frozen embedding at R^2 0.9922. The pathology is the method's, not one reimplementation's. It is present in the authors' released weights, and across four checkpoints long-horizon success rank-orders exactly with metric quality and inversely with prediction accuracy. Replacing only the objective, with nothing retrained and no GPU, lifts goals reached at offset 100 from 26.0% to 98.0%, equals the 98.0% at offset 25, and reaches 92.0% under a third of the budget: planning stops depending on the horizon. The best cost is not the most accurate. A head learned from frame separation alone predicts spatial distance worse than a position probe (r = 0.819 against 0.9897) yet plans better, charging 24% more to cross the environment's dividing wall where squared latent distance charges 4% less. It has learned reachability, not proximity.
We introduce Adaptive LeWorldModel (ALeWM), a world model based on a joint-embedding predictive architecture (JEPA) that learns to concentrate predictive information in compact prefixes of a wide latent representation. To encourage this ordering, ALeWM learns a sequence-conditioned distribution over prefix lengths and trains the predictor to estimate the full next embedding from a sampled input prefix. As standard anti-collapse objectives encourage variation across latent coordinates and do not organize them by predictive importance, we also introduce MixSIGReg. MixSIGReg regularizes the masked embeddings against a prior-weighted mixture with Gaussian active prefixes and zeros in the remaining coordinates. As a result, the ALeWM objective encourages early coordinates to retain information useful for prediction and recursive planning. Our analysis shows that the mixture distribution used by MixSIGReg assigns higher variance to earlier coordinate blocks and lower variance to later ones. In addition, we show that, under specified assumptions, prediction error is minimized by placing the information most useful for prediction in earlier blocks. Empirically, we study the behavior of ALeWM in a controlled dynamical system with known state variables and in goal-conditioned visual control. We show that ALeWM consistently achieves higher mean success rates than tuned fixed-width LeWM, with lower planning capacity on average.
A latent world model may achieve accurate short-horizon prediction while still inducing a latent space that is poorly aligned with planning. A key issue is spatiotemporal mismatch: these models are often trained with local predictive supervision, but deployed for long-horizon goal-directed search in latent spaces where Euclidean distance may not reflect what is reachable within a finite action budget. We present the Reachability-Correction auxiliary objective (RC-aux), a lightweight correction for this mismatch in reconstruction-free latent world models. RC-aux keeps the world-model backbone unchanged and adds planning-aligned supervision along two axes. Along the time axis, multi-horizon open-loop prediction trains the model beyond one-step consistency. Along the space axis, budget-conditioned reachability supervision, together with temporal hard negatives, encourages the latent space to distinguish states that are eventually reachable from those reachable within the current planning horizon. At test time, the learned reachability signal can also be used by a reachability-aware planner to favor trajectories that are both goal-directed and attainable under the available budget. We instantiate RC-aux on LeWorldModel and evaluate it under both continuation-training and matched-from-scratch settings. Across goal-conditioned pixel-control tasks and a LIBERO-Goal extension, RC-aux improves LeWM-style planning with modest additional cost. These results suggest that planning with latent world models depends not only on predictive accuracy, but also on whether the learned representation encodes the temporal and geometric structure required by downstream search. The code is available at https://github.com/Guang000/RC-aux.