Latent Geometry Beyond Search: Amortizing Planning in World Models

arXiv 2026 57.4 method

TLDR

Amortizes planning in world models by learning a latent inverse-dynamics mapping, replacing online search with a lightweight model that matches CEM performance at 100-130x lower cost.

Reasoning

The paper presents a novel method for amortizing planning in world models, with strong empirical results across multiple benchmarks. However, it lacks real-world experiments and does not address generative or video world models directly, limiting the scope of its claims.

Read-first score

Read-first score 57.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 25.

Recency 6%

100

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

Reproducibility 18%

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

Methodology quality 18%

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

Citation impact 18%

71.7

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.71658155

Topical relevance 29%

35.7

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

Citation velocity 12%

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

Field roles

FrontierBridgeMethodology anchorReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 572.

Keyword Scores

world model

model-based reinforcement learning world model

world dynamics prediction

generative world model

world simulator

interactive world model

video world model

Deep Analysis

Innovations

Amortizing planning into a latent inverse-dynamics mapping instead of requiring online search
Goal-Conditioned Inverse Dynamics Model (GC-IDM) that maps current latent state, goal latent state, and remaining horizon directly to the next action
Demonstrating that sufficiently structured latent spaces can shift part of the planning burden from online optimization to learned inference

Methodology

The study uses a pretrained LeWorldModel whose latent geometry is regularized for smoothness and uniformity. Instead of iterative planning, they introduce a lightweight Goal-Conditioned Inverse Dynamics Model (GC-IDM) that takes the current latent state, goal latent state, and remaining horizon as input and outputs the next action. Evaluation is conducted across four benchmark environments (navigation, contact-rich manipulation, continuous control) against test-time planners including CEM, MPPI, iCEM, and gradient-based methods.

Key Results

The GC-IDM matches or exceeds CEM in seven of eight environment-protocol settings while reducing per-decision cost by 100-130x. A broader sweep over test-time planners confirms that this result is not specific to a particular optimizer.

Limitations

The approach relies on a pretrained world model with specific latent geometry regularization (smoothness and uniformity), which may not generalize to other world models or latent spaces
Evaluation is limited to four benchmark environments; scalability to more complex or diverse tasks is not demonstrated
The amortization may fail if the latent representation does not locally encode the structure needed for planning

Resources

Paper Code