I2 -World: Intra-Inter Tokenization for Efficient Dynamic 4D Scene Forecasting

arXiv 25.7 2025 63.2 method

TLDR

Proposes I2-World, an efficient encoder-decoder framework with intra-inter tokenization for 4D occupancy forecasting in autonomous driving.

Reasoning

Strengths include a novel decoupled tokenization approach, significant performance improvements (25.1% mIoU, 36.9% IoU), and computational efficiency. Weaknesses are the lack of explicit dataset details and discussion of limitations in the abstract.

Read-first score

Read-first score 63.2, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 39.

Recency 8%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Reproducibility 25%

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

Topical relevance 42%

55.7

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

Methodology quality 25%

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

Field roles

FrontierReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 252.

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

Decoupling scene tokenization into intra-scene and inter-scene tokenizers
Intra-scene tokenizer using multi-scale residual quantization for hierarchical compression preserving spatial details
Inter-scene tokenizer residually aggregating temporal dependencies across timesteps
Encoder-decoder architecture instead of decoder-only GPT-style autoregressive models
Encoder aggregates spatial context and predicts a transformation matrix for high-level control over scene generation
Decoder conditioned on transformation matrix and historical tokens ensures temporal consistency

Methodology

I2-World proposes a dual tokenization approach: an intra-scene tokenizer using multi-scale residual quantization to hierarchically compress 3D scenes while preserving spatial details, and an inter-scene tokenizer that residually aggregates temporal dependencies across timesteps. The model adopts an encoder-decoder architecture where the encoder aggregates spatial context from the current scene and predicts a transformation matrix for high-level control, and the decoder, conditioned on this matrix and historical tokens, ensures temporal consistency during generation.

Key Results

I2-World achieves state-of-the-art performance, outperforming existing methods by 25.1% in mIoU and 36.9% in IoU for 4D occupancy forecasting, while requiring only 2.9 GB of training memory and achieving real-time inference at 37.0 FPS.

Resources

Paper Code