New: the Timeline is live. Track world model releases, papers, and benchmark updates in real time.
world-models.io
The Knowledge Hub for AI World Models

LeWorldModel

LeWorldModel (LeWM) is a compact 15M-parameter JEPA architecture that learns physical reality on a single GPU, planning up to 48x faster than foundation models. Introduces SigReg to solve the JEPA collapse problem.

robotics model-based-rl simulation embodied-ai

Key Attributes

Static HTML snapshot of the model record for crawlers and no-JS readers.

AttributeValue
ModelLeWorldModel
Lab / OrganizationMila / NYU / Samsung SAIL
CategorySelf-Supervised World Model
SubtypeCompact JEPA Model
World Model TypeCompact joint-embedding predictive world model for physical understanding
Primary DomainPhysical Reasoning
ArchitectureCompact Vision Transformer with JEPA and SigReg regularization
ModalityVisual → Latent Predictions
Training MethodSelf-supervised JEPA training with SigReg collapse prevention, single-GPU trainable
Statusactive
Year2026
Performance Index78/100 (medium confidence, v1.1)

About LeWorldModel

Main editorial body preserved directly in static HTML.

LeWorldModel (LeWM) is a compact JEPA-based world model with only 15 million parameters that achieves physical understanding and planning up to 48× faster than foundation-model-based alternatives. Developed by researchers from Mila, NYU, and Samsung SAIL, it introduces SigReg, a novel regularization technique that solves the representation collapse problem that has plagued JEPA architectures. LeWM demonstrates that large-scale foundation models are not strictly necessary for physical reasoning: efficient, well-regularized architectures can learn meaningful world representations from pixels on commodity hardware.

LeWorldModel is a compact joint-embedding predictive world model for physical understanding developed by Mila / NYU / Samsung SAIL in 2026 for physical reasoning.

Editorial Snapshot

Short extractable facts for answer engines and no-JS readers.

SignalValue
DefinitionLeWorldModel is a compact joint-embedding predictive world model for physical understanding developed by Mila / NYU / Samsung SAIL in 2026 for physical reasoning.
Short DescriptionA compact 15M-parameter JEPA world model that learns real-world physics on a single GPU, solving the notorious representation collapse problem.
Benchmark Rows2
FAQ Entries1
Related Models4
Related Guides0
Related Research Topics0
Last Updated2026-04-10

Notable Features

Key capabilities associated with this model.

  • Only 15M parameters
  • Trainable on a single GPU
  • SigReg solves JEPA collapse problem
  • 48× faster planning than foundation models
  • End-to-end from pixels

Use Cases

Representative applications attached to this model record.

Physical scene understandingEfficient robot planningResource-constrained deploymentResearch prototyping

Strengths and Limitations

Balanced assessment surfaced in static HTML.

Strengths

  • Extremely compact
  • Single-GPU training
  • Solves collapse problem
  • Fast inference
  • Open research

Limitations

  • Limited to simple physical scenarios
  • Early-stage research
  • Not yet tested at scale

Benchmarks

Published benchmark evidence attached to this model record.

BenchmarkMetricResultSource
Planning EfficiencyPlanning Speedup 48 xUp to 48x fasterSource
Training FootprintHardware Requirement 1 GPUSingle-GPU trainableSource

References and Citations

Primary references preserved in static HTML for citation extraction.

ReferenceLink
LeCun et al., 2026. LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels. arXiv:2603.19312.Open source

Related Models

Nearby models linked from the current editorial record.

ModelCategoryWorld Model TypeIndex v1.1
V-JEPASelf-Supervised World ModelSelf-supervised visual world model70/100
V-JEPA 2Self-Supervised World ModelJoint-embedding predictive world model for video understanding and robot planning87/100
I-JEPASelf-Supervised World ModelSelf-supervised visual world model61/100
AMI World ModelFoundation World ModelMultimodal generative world model38/100

Direct Comparisons

Side-by-side comparisons already connected to this model.

ComparisonMatchupSummary
V-JEPA 2 vs V-JEPAV-JEPA 2 vs V-JEPAV-JEPA 2 dramatically scales up Meta FAIR's self-supervised video world model, achieving state-of-the-art visual understanding and zero-shot robot control, capabilities V-JEPA didn't demonstrate.
V-JEPA 2 vs I-JEPAV-JEPA 2 vs I-JEPATwo milestones of the JEPA roadmap: I-JEPA established self-supervised image representation by predicting in latent space; V-JEPA 2 extends the paradigm to video at foundation scale and demonstrates zero-shot robot control.
LeWorldModel vs DreamerV3LeWorldModel vs DreamerV3LeWorldModel revisits LeCun's energy-based JEPA philosophy for control, predicting in latent space without pixel reconstruction. DreamerV3 remains the canonical RSSM-based agent that learns by imagining pixel-grounded rollouts.
Genie 3 vs V-JEPA 2Genie 3 vs V-JEPA 2Two green-index leaders that represent different frontier philosophies. Genie 3 is an interactive generative world model that turns text into playable environments, while V-JEPA 2 is a self-supervised latent predictor optimized for physical reasoning and zero-shot robot planning.

Frequently Asked Questions

FAQ answers rendered directly into static HTML for extractable responses.

What is SigReg?

SigReg is a regularization technique introduced in LeWorldModel that prevents the representation collapse problem in JEPA architectures by constraining the singular values of the latent representations, ensuring they remain informative throughout training.

Quick Answer

Short extractable summary preserved directly in static HTML.

  • LeWorldModel is a compact joint-embedding predictive world model for physical understanding developed by Mila / NYU / Samsung SAIL in 2026 for physical reasoning.
  • Use this page when you need a fast read on how LeWorldModel fits into the self-supervised world model landscape, then validate the details in the benchmarks, citations, and related pages.
  • A key strength surfaced in the editorial record is extremely compact.

Editorial Trust Signals

Editorial provenance and refresh policy preserved directly in static HTML.

Published by world-models.io editorial board.

Lead editor Tyler D. - Technical editor, methodology and benchmark analysis.

This model page synthesizes primary papers, official model pages, benchmark evidence, and related world-models.io context into a reference resource.

Each editorial page is assembled from primary sources, normalized into extractable summaries, checked for factual drift, and reviewed before publication or major refreshes. Last reviewed: 2026-04-10.

Pages are refreshed when a new paper, benchmark, release, architecture update, or stronger primary source materially changes the answer a reader or AI system should retrieve.

Each page links back to relevant primary sources and keeps a stable canonical URL so readers can verify claims, trace context, and reference the most up-to-date version. See the editorial policy.

Primary sources onlyLast reviewed date visibleMethodology documentedSource links included

External Sources

Primary model and lab sources embedded in static HTML.

References

  1. [1] LeCun et al., 2026. LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels. arXiv:2603.19312.