Emu3.5: Native Multimodal Models are World Learners

arXiv 25.10 2025 57.8 method

TLDR

Emu3.5 is a large-scale multimodal world model trained on internet videos with next-token prediction, achieving strong generation and world-modeling abilities.

Reasoning

Strengths include a unified next-token prediction framework across vision and language, efficient inference via DiDA, and demonstrated world-modeling capabilities. Weaknesses are limited details on real-world embodied manipulation evaluation and comparison only to one model.

Read-first score

Read-first score 57.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 48.

Recency 8%

86.7

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

Topical relevance 42%

68.6

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=result

Reproducibility 25%

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 369.

Keyword Scores

world model

world dynamics prediction

generative world model

video world model

interactive world model

world simulator

model-based reinforcement learning world model

Deep Analysis

Innovations

Unified next-token prediction across vision and language for native multimodal world modeling
Discrete Diffusion Adaptation (DiDA) for converting token-by-token decoding into bidirectional parallel prediction, achieving ~20x inference speedup
Large-scale reinforcement learning post-training to enhance multimodal reasoning and generation
Long-horizon vision-language generation, any-to-image (X2I) generation, and complex text-rich image generation
Spatiotemporally consistent world exploration and open-world embodied manipulation across diverse scenarios

Methodology

Emu3.5 is pre-trained end-to-end with a unified next-token prediction objective on a corpus of vision-language interleaved data containing over 10 trillion tokens, primarily derived from sequential frames and transcripts of internet videos. It is further post-trained with large-scale reinforcement learning to enhance multimodal reasoning and generation. To improve inference efficiency, Discrete Diffusion Adaptation (DiDA) converts token-by-token decoding into bidirectional parallel prediction.

Key Results

Emu3.5 achieves performance comparable to Gemini 2.5 Flash Image on image generation and editing tasks and demonstrates superior results on a suite of interleaved generation tasks. DiDA accelerates per-image inference by about 20x without sacrificing performance.

Resources

Paper