World Model Self-Distillation: Training World Models to Solve General Tasks

arXiv 2026 64.5 method

TLDR

Proposes self-distillation and RL to train video diffusion world models for task-solving without curated task-video supervision, outperforming on benchmarks.

Reasoning

The paper presents a scalable framework combining self-distillation and reinforcement learning to elicit task-solving abilities from pretrained video generators, avoiding costly supervised data. Strengths include novel use of VLM feedback and demonstrated transfer to robotics; weaknesses include reliance on VLM for task generation and evaluation, which may introduce biases.

Read-first score

Read-first score 64.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 49.

Recency 6%

100

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

Citation impact 18%

95.7

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.95741109

Methodology quality 18%

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

Topical relevance 29%

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=none

Citation velocity 12%

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

Field roles

FoundationFrontierBridgeMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 419.

Keyword Scores

world model

video world model

generative world model

world dynamics prediction

world simulator

model-based reinforcement learning world model

interactive world model

Deep Analysis

Innovations

Combining self-distillation with reinforcement learning to elicit task-solving abilities in pretrained video diffusion models without curated task-video supervision.
Using a vision-language model to automatically generate candidate tasks and detailed step-by-step solutions, enabling scalable training.
Distilling the Demonstrator (caption-guided video diffusion) into an Executor conditioned only on an image and a short task prompt, transferring execution knowledge.
Reinforcement learning from VLM feedback to further improve the Executor, exploiting the asymmetry between judging and generating task solutions.

Methodology

The framework uses a vision-language model (VLM) to generate a candidate task and a detailed step-by-step solution from an unlabeled scene image. This solution conditions a pretrained video diffusion model (the Demonstrator), and its behavior is distilled into an Executor that is conditioned only on the image and a short task prompt. The Executor is then further improved via reinforcement learning from VLM feedback, where the VLM judges whether a sampled video satisfies the task.

Key Results

On the proposed WorldTasks-Benchmark and the DreamGen robotics benchmark, the Executor surpasses the Demonstrator under a VLM-based evaluation protocol and transfers competitively to robotic tasks.

Limitations

Reliance on a vision-language model for task generation and evaluation may introduce biases or inaccuracies in task definitions and judgments.
The approach requires a pretrained video diffusion model and a VLM, which may not be available or optimal for all domains.
Evaluation is based on a VLM-based protocol, which may not fully align with real-world task success metrics.
The framework is limited to tasks that can be described textually and solved via video generation, potentially excluding tasks requiring physical interaction or precise control.

Resources

Paper