Skip to content

Render, Don't Decode: Weight-Space World Models with Latent Structural Disentanglement

Home / Papers / Render, Don't Decode: Weight-Space World Models with Latent Structural Disentanglement

Render, Don’t Decode: Weight-Space World Models with Latent Structural Disentanglement

Authors: Roussel Desmond Nzoyem, Mauro Comi Date: 2026-05-07 Paper ID: openalex:2605.06298

Summary

NOVA is a world modeling framework that replaces traditional decoder-based architectures with an analytically rendered weight-space representation using coordinate-based implicit neural representations (INRs). By operating directly on model weights rather than latent vectors, the framework achieves high computational efficiency, portability, and zero-shot super-resolution. The approach inherently disentangles complex scene dynamics into structural components like background and foreground, enabling high-fidelity controllable video forecasting.

Key Contributions

  • Introduces NOVA, a world model that encodes state as weights of an implicit neural representation, eliminating the need for heavy decoders.
  • Achieves emergent latent disentanglement of scene components like foreground, background, and motion without auxiliary supervision.
  • Demonstrates efficient controllable video forecasting on a single consumer GPU at ~40M parameters.

Key Concepts

  • weight-space-world-models: A world modeling framework that encodes environment states into the weights and biases of coordinate-based implicit neural representations.

Archivist Review

The proposed ‘Weight-Space World Models’ concept is highly novel and represents a significant departure from standard latent state modeling in world models, making it a valuable addition to the vault. I rejected ‘Latent Structural Disentanglement’ because the paper frames it as an emergent capability of the weight-space architecture rather than a distinct, modular, or independently reusable mechanism. I applied a restrictive filter to ensure only the central paradigm shift was captured.

Approved Concepts

  • Weight-Space World Models: This introduces a paradigm shift in world modeling by encoding environment states directly into the parameters of a coordinate-based neural network rather than using traditional latent vector spaces, effectively bypassing the decoder bottleneck.

Rejected Candidates

  • [concept] Latent Structural Disentanglement (latent-structural-disentanglement) - subcomponent_of_broader_mechanism: While mentioned as an outcome of the model, it is an emergent property rather than a distinct, standardized architectural mechanism.

Metadata & Links

url
https://arxiv.org/abs/2605.06298
paper_id
2605.06298
paper_source
openalex
domain
computer-vision
tags
architectures
datasets
concept_slugs
weight-space-world-models
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-05-10T06:06:03Z