Enfold: Folding World Model Imagination into Predictive Representations for Ultra-Efficient Embodied Control

arXiv 2026 42.4 method, application

TLDR

Enfold transfers world generative model computation into a representation predicted from current context, enabling efficient embodied control without executing the generator at deployment.

Reasoning

The paper presents a novel approach to reuse the internal computation of world generative models for representation learning, showing strong empirical results with significant latency reductions. However, the abstract lacks methodological details and explicit limitations, though the real-robot experiments strengthen its claims.

Read-first score

Read-first score 42.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 50.

Recency 6%

100

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

Topical relevance 29%

71.4

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

Methodology quality 18%

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

Reproducibility 18%

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

Citation impact 18%

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

Citation velocity 12%

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 379.

Keyword Scores

world model

generative world model

world dynamics prediction

video world model

interactive world model

model-based reinforcement learning world model

world simulator

Deep Analysis

Innovations

Internalizing future-generative computation into a present-only representation by distilling multi-level intermediate states of a world generator processing the observed future.
Training a current-only encoder supervised by generator states, with the representation fed back to condition future generation and read by task heads without allowing task gradients to reshape the encoder.
Eliminating the generator at deployment to achieve 3.7× (Enfold) and 10.1× (Enfold-Flash) action latency reduction while maintaining strong control performance.
Learning representations that suppress nuisance variation and preferentially capture changes that emerge over longer horizons.
Demonstrating adaptive behavior under human interventions where both the generated continuation and executed actions change, inconsistent with fixed trajectory replay.

Methodology

Enfold trains a current-only encoder (from visual context and language instruction) to match the multi-level intermediate states of a world generator that processes the observed future sequence. The learned representation conditions future generation and is used by task heads for action prediction, but task gradients do not update the encoder. At test time, action prediction uses only the encoder without executing the generator, enabling fast inference.

Key Results

Enfold achieves strong control on LIBERO, RoboTwin2.0, and real-robot tasks, reducing action latency by 3.7× relative to Fast-WAM (Enfold) and 10.1× (Enfold-Flash); the representation suppresses nuisance variation, captures long-horizon changes, and adapts both generated continuations and actions when the current scene is altered by human intervention.

Resources

Paper