AnyGoal: On World Model Generalization Beyond Expert Rollouts

arXiv 2026 40.6 benchmark

TLDR

Introduces AnyGoal benchmark to test compositional generalization of world models beyond expert rollouts, finding current models fail on recombined state-action factors.

Reasoning

The paper provides a clear benchmark and systematic analysis of world model generalization, with a useful open-source library. However, the abstract lacks quantitative results and concrete limitations, making the empirical impact less visible at this level.

Read-first score

Read-first score 40.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 47.

Recency 5%

100

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

Methodology quality 15%

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

Topical relevance 25%

67.1

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

Reproducibility 15%

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

Graph prestige 15%

5.7

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

Citation impact 15%

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

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: 555.

Keyword Scores

world model

world dynamics prediction

world simulator

model-based reinforcement learning world model

generative world model

interactive world model

video world model

Deep Analysis

Innovations

Introduces AnyGoal, a benchmark of goal-shifted task variants built from widely used world-model environments, together with an open-source library.
Frames world model generalization beyond expert rollouts as compositional generalization over familiar state and action factors.
Provides a systematic analysis of dynamics and representation failures across model families, interaction regimes, and evaluation metrics.

Methodology

The paper introduces AnyGoal, a benchmark of goal-shifted task variants built from widely used world-model environments. Each task keeps individual state factors and actions within their training support but makes their combinations underrepresented in expert trajectories. The authors evaluate multiple world model families across environments and examine dynamics and representation failures across interaction regimes and metrics.

Key Results

Across multiple environments and world model families, current models consistently exhibit failures under recombinations of familiar state and action factors. The analysis shows where current world models generalize and where challenges remain for goal-shifted prediction and planning.

Limitations

Current world models consistently fail under recombinations of familiar state and action factors.
Challenges remain for goal-shifted prediction and planning.
The relationship between data coverage, dynamics generalization, and downstream planning performance remains underexplored.

Resources

Paper