Driving · 2026-11-13
The Posterior-Prior Epistemic Gap: When World Model Imagination Fails to Carry Task-Relevant Signals
Jaerock Kwon, Elahe Delavari, Donghyun Kim, Haewoon Nam
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
The authors report a measured posterior-prior epistemic gap in an RSSM world model with active-inference planning for CARLA driving: an obstacle probe scores AUC 0.85 on observation-filtered posterior states, but decays on imagined prior states to AUC 0.71 after 15 imagination steps (57% logit-separation decay). Obstacle avoidance stays at 0% without privileged runtime coordinates and ensemble epistemic disagreement fails to flag the loss.
Why it matters
Model-based planners often assume imagined rollouts preserve task-relevant signals, but the authors directly test this by probing the same auxiliary obstacle head on posterior vs imagined prior states. They also show that the planner’s ensemble-based epistemic term can remain at chance-level obstacle discrimination despite the imagination signal loss, creating an unflagged “unknown unknown” failure mode at planning time.
Method
- Use an RSSM world model with active-inference planning in CARLA, distinguishing posterior (observation-filtered) vs prior (imagined) latent states during rollouts.
- Train a supervised obstacle probe/auxiliary head on posterior states and apply the same frozen head to imagined prior states at each imagination depth to measure task-relevant signal decay.
- Test whether ensemble-based epistemic disagreement flags the gap by using ensemble disagreement as a binary obstacle classifier across imagination depths.
Limitation
This is a single-seed diagnostic case study: one trained world model and limited evaluation (11 episodes across 3 fixed highway routes with static obstacles). The auxiliary head is trained on posterior states and evaluated on prior states, introducing a distribution-shift confound; the probe retrained on prior states is stated as a clean next test but not completed in the setup.
Abstract (from arXiv)
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 \textbf{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$\rightarrow$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 \textbf{task-conditioned weighting of Effective Free Energy components} over the same world model yields lane keeping ($81.2 \pm 1.8$\% route completion) and obstacle avoidance (28/28 avoided, \emph{using privileged coordinates}) in the CARLA simulator, with runtime weight modulation enabling post-evasion centering.
Related papers
- The Planning Limits of Latent World Models
- Completion Aware Guidance for World Action Models
- Second Thoughts: Exploring and Planning through Comparative Imagination in World Models
- Order-Consistent Latent Geometry for Reliable Short-Horizon Planning in World Models
- When Reality Outpaces Cognition: A New Scaling Principle for Reliable Intelligence
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First announced by arXiv: 2026-11-13 · Indexed by Pub-AI: 2026-11-13 UTC
The Posterior-Prior Epistemic Gap: When World Model Imagination Fails to Carry Task-Relevant Signals
First announced by arXiv: 2026-11-13 · Original record ↗
- arxiv
- 2611.14955
Explicit source versions (1)
- v1 · Date not supplied
- Preprint available
Source date: 2026-11-13 · First detected: 2026-11-13 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · RSS guid=oai:arXiv.org:2611.14955v1; arxiv:announce_type=new; item pubDate (official metadata)
Abstract · arxiv
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 \textbf{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$\rightarrow$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 \textbf{task-conditioned weighting of Effective Free Energy components} over the same world model yields lane keeping ($81.2 \pm 1.8$\% route completion) and obstacle avoidance (28/28 avoided, \emph{using privileged coordinates}) in the CARLA simulator, with runtime weight modulation enabling post-evasion centering.
