World2VLM: Distilling World Model Imagination into VLMs for Dynamic Spatial Reasoning

arXiv 2026 54.3 method

TLDR

Distills a generative world model into a VLM via synthetic future views to improve dynamic spatial reasoning, outperforming test-time coupling.

Reasoning

The paper presents a novel training framework that leverages a world model to generate structured supervision for spatial reasoning, achieving consistent benchmark improvements without inference-time overhead. Strengths include a clear methodology and empirical validation on multiple benchmarks; weaknesses include limited discussion of generalization beyond camera trajectories and potential reliance on synthetic data quality.

Read-first score

Read-first score 54.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 38.

Recency 6%

100

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

Methodology quality 18%

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

Citation impact 18%

66.1

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.66058576

Topical relevance 29%

54.3

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 18%

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

Citation velocity 12%

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

Field roles

FrontierBridgeMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 319.

Keyword Scores

world model

generative world model

world dynamics prediction

world simulator

video world model

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Distilling spatial imagination from a generative world model into a vision-language model via a training framework
Using a view-consistent world model to synthesize geometrically aligned future views for structured forward and inverse spatial reasoning supervision
Two-stage post-training recipe on a compact dataset generated by the pipeline
Demonstrating that world models can serve as effective training-time teachers, not just inference-time tools

Methodology

World2VLM uses a view-consistent world model to synthesize geometrically aligned future views given an initial observation and a parameterized camera trajectory. It derives structured supervision for both forward (action-to-outcome) and inverse (outcome-to-action) spatial reasoning. The VLM is post-trained with a two-stage recipe on a compact dataset generated by this pipeline, and evaluated on multiple spatial reasoning benchmarks including SAT-Real, SAT-Synthesized, VSI-Bench, and MindCube.

Key Results

World2VLM delivers consistent improvements over the base model across diverse benchmarks and outperforms test-time world-model-coupled methods while eliminating the need for expensive inference-time generation.

Resources

Paper