Latent Dynamics · 2026-10-30
Structural-Test Geometry as an Inductive Bias for Partially Observable World Models
Todd Y. Zhou
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 propose posterior-marginalized structural tests that induce a pseudometric over histories, then train quotient-preserving (QP) world models by aligning latent Euclidean distances to that metric. Across a fixed six-task diagnostic suite, QP improves planning success by 26.1 points over metric decoding and 6.4 over structured belief distillation under held-out structural variation.
Why it matters
They study when predictive world-model objectives should learn distinctions needed for planning under partial observability. By comparing test-response regression (TR), metric decoding (MD), QP distance alignment, and structured belief distillation (BD-S) under common supervision, the paper shows that accurate relational recovery via decoding can still leave native latent geometry and shift performance worse than direct geometric alignment.
Method
- Define structural tests that restart a privileged simulator from posterior beliefs over histories, optionally apply structural probes, and compare response distributions to form a test pseudometric and quotient geometry.
- Train quotient-preserving (QP) recurrent RSSM models by aligning shared latent Euclidean distances to the held-out structural-test distance targets (distance-scale matching with fitted scale).
- Compare QP against TR (predict response distributions), MD (decode test metric from latent pair decoder without constraining native geometry), and BD-S (privileged task-relevant posterior distillation), discarding auxiliary branches at deployment.
Limitation
The authors condition on a supplied, task-informed structural vocabulary for test construction, stating that automatic test discovery is outside scope.
Abstract (from arXiv)
Predictive world-model objectives need not organize latent histories according to distinctions required for later control. With privileged simulator access during training, should response-derived structural information be predicted, decoded, or imposed directly on latent geometry--and how do these choices compare with structured belief distillation? We define posterior-marginalized structural tests whose response distributions induce a pseudometric over histories, then train quotient-preserving (QP) world models by aligning shared latent Euclidean distances to that metric. Across a fixed six-task diagnostic suite, the largest differences appear under held-out structural variation: QP improves planning success by 26.1 points $[21.5,30.7]$ over metric decoding and 6.4 points $[3.8,9.0]$ over structured belief distillation. On GridWorld, metric decoding accurately recovers held-out structural distances ($\rho=.91$) while its native latent geometry remains more distorted than QP's (paired IQD difference $.22$ $[.20,.24]$), showing that relational recoverability and direct geometric alignment can diverge.
Related papers
- ATLAS: Aligned Transport of Latent Structure for Reliable World Model Planning
- Understanding the limits of latent world models
- When Low Prediction Error Misleads Planning: Diagnosing Representation, Dynamics, and Decision Failures in Latent World Models
- Roz: A Taste Test for World Model Latents
- Order-Consistent Latent Geometry for Reliable Short-Horizon Planning in World Models
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First announced by arXiv: 2026-10-30 · Indexed by Pub-AI: 2026-10-30 UTC
Structural-Test Geometry as an Inductive Bias for Partially Observable World Models
First announced by arXiv: 2026-10-30 · Original record ↗
- arxiv
- 2610.38903
Explicit source versions (1)
- v1 · Date not supplied
- Preprint available
Source date: 2026-10-30 · First detected: 2026-10-30 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · RSS guid=oai:arXiv.org:2610.38903v1; arxiv:announce_type=new; item pubDate (official metadata)
Abstract · arxiv
Predictive world-model objectives need not organize latent histories according to distinctions required for later control. With privileged simulator access during training, should response-derived structural information be predicted, decoded, or imposed directly on latent geometry--and how do these choices compare with structured belief distillation? We define posterior-marginalized structural tests whose response distributions induce a pseudometric over histories, then train quotient-preserving (QP) world models by aligning shared latent Euclidean distances to that metric. Across a fixed six-task diagnostic suite, the largest differences appear under held-out structural variation: QP improves planning success by 26.1 points $[21.5,30.7]$ over metric decoding and 6.4 points $[3.8,9.0]$ over structured belief distillation. On GridWorld, metric decoding accurately recovers held-out structural distances ($ρ=.91$) while its native latent geometry remains more distorted than QP's (paired IQD difference $.22$ $[.20,.24]$), showing that relational recoverability and direct geometric alignment can diverge.
