Co-Evolving Latent Action World Models

arXiv 25.10 2025 66 method

TLDR

CoLA-World jointly trains a latent action model and pretrained video world model via warm-up, enabling co-evolution for better simulation and planning.

Reasoning

The paper introduces a novel joint training paradigm that resolves representational collapse, achieving co-adaptation between the latent action model and world model. Strengths include empirical improvements over two-stage methods, but the abstract lacks specific benchmark details and potential reliance on pretrained models.

Read-first score

Read-first score 66, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 61.

Topical relevance 42%

87.1

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

Recency 8%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Methodology quality 25%

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

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 380.

Keyword Scores

world model

world simulator

generative world model

interactive world model

video world model

world dynamics prediction

model-based reinforcement learning world model

Deep Analysis

Innovations

First successful joint training of latent action model (LAM) and pretrained world model, replacing the two-stage paradigm with a co-evolution cycle.
Critical warm-up phase that aligns representations of the from-scratch LAM with the pretrained world model, preventing representational collapse.
Synergistic co-adaptation where the world model acts as a tutor to shape the LAM, and the LAM provides a more precise control interface to the world model.

Methodology

CoLA-World jointly trains a latent action model (LAM) and a pretrained video generation world model. It employs a critical warm-up phase to align the representations of the randomly initialized LAM with the frozen world model, after which both models are trained jointly in a co-evolution loop. The world model provides gradients to refine the LAM, while the LAM offers a more adaptable control interface for the world model.

Key Results

CoLA-World matches or outperforms prior two-stage methods in both video simulation quality and downstream visual planning tasks, establishing a robust and efficient new paradigm.

Limitations

Dependence on a pretrained video generation world model, which may limit applicability to domains without such models.
Requires a carefully designed warm-up phase to avoid representational collapse, potentially adding tuning complexity.

Resources

Paper