WorldAgen: Unified State-Action Prediction with Test-Time World Model Training

arXiv 2026 43.9 method

TLDR

WorldAgen combines world modeling and action prediction in a VLA transformer with test-time training, adapting to new dynamics and improving performance on CALVIN and LIBERO benchmarks.

Reasoning

The paper proposes a clear unified architecture with separate world-model and agent heads, supported by test-time adaptation and benchmark evaluations. Weaknesses include limited evidence of real-world deployment and no details on scalability or failure cases in the abstract.

Read-first score

Read-first score 43.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 43.

Methodology quality 15%

100

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

Recency 5%

100

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

Topical relevance 25%

61.4

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

Reproducibility 15%

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

Citation impact 15%

13.4

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.13388351

Graph prestige 15%

5.7

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0016

Citation velocity 10%

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

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 682.

Keyword Scores

world model

world dynamics prediction

interactive world model

world simulator

generative world model

model-based reinforcement learning world model

video world model

Deep Analysis

Innovations

Unified framework that jointly learns world modeling and action prediction in a shared Transformer backbone.
Two-head design: a world model head predicts future states from past state-action trajectories, and an agent model head predicts actions conditioned on task instructions.
Mixed Unidirectional Attention Mask to separate the world model and agent model computations.
Test-Time Training (TTT) adaptation mechanism that samples exploratory actions, collects ground-truth state transitions, and applies lightweight updates to the world model at deployment.

Methodology

WorldAgen uses a shared Transformer backbone with two heads: a world model head that predicts future states from past state-action trajectories and an agent model head that predicts actions conditioned on task instructions, separated by a Mixed Unidirectional Attention Mask. During test time, it samples exploratory actions, collects ground-truth state transitions, and performs lightweight Test-Time Training updates to refine the world model. Evaluation is conducted on the CALVIN and LIBERO benchmarks against current state-of-the-art approaches.

Key Results

The baseline WorldAgen model achieves comparable, and in some cases superior, performance to current state-of-the-art approaches on CALVIN and LIBERO. With Test-Time Training on a small number of samples, WorldAgen surpasses existing state-of-the-art models.

Resources

Paper