Learning Intrinsic Dynamics in Latent World Models

arXiv 2026 35.3 method, theory

TLDR

Proposes that action predictability in latent world models preserves intrinsic dynamics, supported by theory and Push-T experiments.

Reasoning

The paper provides a theoretical bound and empirical evidence linking action predictability to preserving control-relevant dynamics, with a simple method. However, evaluation is limited to a simulated Push-T benchmark and the abstract truncates before full results.

Read-first score

Read-first score 35.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 33.

Recency 5%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 15%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=analysis,evaluation,metric,result

Topical relevance 25%

47.1

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 15%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=code

Graph prestige 15%

5.7

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0016

Citation impact 15%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 10%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 460.

Keyword Scores

world dynamics prediction

world model

model-based reinforcement learning world model

interactive world model

generative world model

world simulator

video world model

Deep Analysis

Innovations

Proposes action predictability, reconstructing actions from consecutive latent states, as an inductive bias for preserving intrinsic dynamics in latent world models.
Provides a theoretical result showing a positive lower bound on squared inverse prediction error when an encoder merges distinguishable outcomes of different actions from the same state and appearance, even while forward prediction can remain exact.
Introduces Concat IDM, which predicts actions from concatenated latent states and updates the shared encoder using action labels already available in offline trajectories without changing the test-time planner.
Empirically validates the role of action prediction through controlled appearance-diversity comparisons, supervision interventions, and latent probes.

Methodology

The authors formalize action predictability as reconstructing an action from consecutive latent states and derive a lower bound on squared inverse prediction error when an encoder merges distinguishable outcomes of different actions from the same state and appearance. They validate the hypothesis using controlled appearance-diversity comparisons, supervision interventions, and latent probes, and instantiate the idea as Concat IDM, which predicts actions from concatenated latent states and updates the shared encoder with action labels from offline trajectories while leaving the test-time planner unchanged. Evaluation uses Push-T with 24 appearances over three independent training runs, with planning success and paired improvement over single-appearance training as metrics.

Key Results

On Push-T with 24 appearances, Concat IDM increases planning success from 9.1% to 77.2% across three independent training runs, and its paired improvement is 52.9 percentage points larger than the improvement observed with a single appearance. The theoretical analysis shows that merging distinguishable outcomes of different actions from the same state and appearance yields a positive lower bound on squared inverse prediction error while forward prediction can still be exact.

Resources

Paper