A Frame is Worth One Token: Efficient Generative World Modeling with Delta Tokens

arXiv 2026 56.9 method

TLDR

Introduces DeltaTok and DeltaWorld, a generative video world model using delta tokens in VFM feature space to efficiently predict diverse futures with fewer parameters and FLOPs.

Reasoning

The paper presents a novel tokenization scheme and generative world model with strong efficiency gains and empirical evaluation on dense forecasting tasks. Weaknesses include lack of visible details on benchmark scope and potential limitations of delta-token representation for broader video generation.

Read-first score

Read-first score 56.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 37.

Recency 8%

100

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

Methodology quality 25%

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

Topical relevance 42%

52.9

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 25%

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 361.

Keyword Scores

generative world model

world model

video world model

world dynamics prediction

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

DeltaTok tokenizer encodes VFM feature differences between consecutive frames into a single continuous delta token
DeltaWorld generative world model operates on delta tokens to efficiently generate diverse plausible futures
Delta tokens reduce video from 3D spatio-temporal representation to 1D temporal sequence, e.g., 1,024x token reduction for 512x512 frames
Enables tractable multi-hypothesis training by generating many futures in parallel and supervising only the best
At inference, produces diverse predictions in a single forward pass

Methodology

The method predicts futures in the feature space of a vision foundation model. DeltaTok encodes VFM feature differences between consecutive frames into single continuous delta tokens, and DeltaWorld is a generative world model trained on these tokens. Training uses multi-hypothesis supervision where many futures are generated in parallel and only the best is supervised, with evaluation on dense forecasting tasks.

Key Results

On dense forecasting tasks, DeltaWorld forecasts futures that more closely align with real-world outcomes while having over 35x fewer parameters and using 2,000x fewer FLOPs than existing generative world models.

Resources

Paper