The Posterior-Prior Epistemic Gap: When World Model Imagination Fails to Carry Task-Relevant Signals
Authors: Jaerock Kwon, Elahe Delavari, Donghyun Kim, Haewoon Nam
Organizations: Department of Electrical and Computer Engineering, University of Michigan–Dearborn · Department of Electrical Engineering, Hanyang University, South Korea
Model-based planners that roll out imagined futures assume the transition model faithfully propagates task-relevant signals---a property rarely measured directly. In a single-seed case study of a Recurrent State Space Model world model with active-inference planning in the CARLA driving simulator, we identify a posterior-prior gap---posterior (observation-filtered) states retain obstacle information that imagined rollout states progressively lose: a supervised obstacle probe achieves AUC 0.85 on posterior states, but the same probe applied to imagined prior states shows monotonic decay (AUC 0.85→0.71 over 15 steps; logit separation drops 57%). Because the probe is identical, the decay isolates the transition model---not the representation---as the information bottleneck. Obstacle avoidance drops to 0% without privileged runtime coordinates, and the planner's own ensemble-based epistemic term does not flag the loss---ensemble disagreement yields AUC 0.52 as an obstacle classifier (chance level), making the failure an unflagged imagination failure---an unknown unknown at planning time. As a secondary contribution, we show that task-conditioned weighting of Effective Free Energy components over the same world model yields lane keeping (81.2±1.8% route completion) and obstacle avoidance (28/28 avoided, using privileged coordinates) in the CARLA simulator, with runtime weight modulation enabling post-evasion centering.
Figures & tables
Parameter
Task A
Task B
βinstrumental
1.0
1.0
βepistemic
0.1
0.1
βstate
0.5
0.0
βobstacle
0.0
40.0
Horizon H
12
15
Table 1: Task-conditioned EFE component weights. Same world model checkpoint; only β weights and horizon differ.
Route
Eps
SR
Avoid
MLD
LC
R0 (373m, 2 obs)
5
100%
10/10
0.67±.05
6.0
R1 (518m, 3 obs)
3
100%
9/9
0.66±.03
6.0
R2 (518m, 3 alt)
3
100%
9/9
0.67±.02
6.3
Total
11
100%
28/28
0.66±.04
6.1
Table 2: Obstacle avoidance results using privileged runtime coordinates (mean ± std over episodes). SR: Success Rate; LC: Lane Changes per episode; MLD: Mean Lateral Deviation (m).
Variant
SR
Avoid
MLD
Compl.
Full system
100%
9/9
0.66
99.2%
w/o privileged
0%
0/9
0.28
43.8%
+ Centering
100%
9/9
0.52
99.1%
Table 3: Ablation on R1 (518m, 3 obstacles, 3 episodes each). Same 64 × 64 model throughout. “w/o privileged” removes all six privileged mechanisms simultaneously (not individually isolated); Compl.: route completion.
Method
Δ
MLD
Avoid
Single GMM (different modalities)
64 × 64 RGB
− 24.6
0.31
0%
128 × 128 RGB
− 2.3
0.17
0%
64 × 64 RGBD
− 68.1
—
0%
Contrastive preference ( α=0.2 )
Without aux head
+1.1
0.17
0%
Table 4: Preference gap Δ across modalities and methods. Top rows vary input modality (separate world models); bottom rows vary the preference objective (same 64 × 64 model). α : contrastive down-weighting factor.
State
AUC
Separation
Posterior
0.851
4.89
Prior k=1
0.788
3.59
Prior k=5
0.768
3.01
Prior k=10
0.742
2.54
Prior k=15
0.707
2.11
Table 5: Obstacle probe on posterior vs. imagined prior states. AUC and logit separation (mean obstacle − mean clean) decay with imagination depth.
Figure 1: Posterior-prior epistemic gap across imagination depth. AUC (left axis) and logit separation (right axis) both decay monotonically as the RSSM rolls out imagined states without observations. AUC drops from 0.85 (posterior) to 0.71 ( k=15 ), with logit separation decaying 57% over 15 steps. The sharpest drop occurs at k=1 .