Скачать или смотреть Emu3.5: Native Multimodal Models are World Learners (Oct 2025)

Title: Emu3.5: Native Multimodal Models are World Learners (Oct 2025)
Link: http://arxiv.org/abs/2510.26583v1
Date: October 2025

Summary:
Emu3.5 is a large-scale multimodal world model designed to natively predict the next state across vision and language. It's pre-trained end-to-end on over 10 trillion vision-language interleaved tokens, primarily from internet videos, using a unified next-token prediction objective. Post-training involves large-scale reinforcement learning to enhance multimodal reasoning and generation. For inference efficiency, the paper introduces Discrete Diffusion Adaptation (DiDA), accelerating per-image inference by about 20x without sacrificing performance. Emu3.5 demonstrates strong native multimodal capabilities, including long-horizon vision-language generation, any-to-image (X2I) generation, complex text-rich image generation, and generalizable world-modeling abilities for spatiotemporally consistent world exploration and embodied manipulation.

Key Topics:
Multimodal World Model
Next-Token Prediction
Vision-Language Interleaved Data
Reinforcement Learning (RL)
Discrete Diffusion Adaptation (DiDA)
Any-to-Image Generation (X2I)
Long-Horizon Generation
World Modeling
Embodied Manipulation
Transformer Architecture

Chapters:
00:00 - EMU 3.5 Overview & Core Ideas
00:52 - Solving Inference Speed with DiDA
01:47 - Model Architecture Deep Dive
02:49 - Efficient Vision Tokenization (IBQ)
04:20 - World Learner Training & Data
06:21 - Post-Training: SFT and RL
08:00 - Tackling Slow Image Generation
08:37 - DiDA: Parallel Refinement & Control
10:29 - Performance: World Models & Narrative
12:26 - Performance: Image Generation & Editing
13:33 - Conclusion & Future Outlook

Stock video credits:
Pixabay - https://www.pexels.com/@pixabay
Colin Jones - https://www.pexels.com/@larchmedia
Charlie Mounsey - https://www.pexels.com/@charlie-mouns...
Pressmaster - https://www.pexels.com/@pressmaster
Pavel Danilyuk - https://www.pexels.com/@pavel-danilyuk
StefWithAnF - https://www.pexels.com/@stefwithanf-1...
Pachon in Motion - https://www.pexels.com/@pachon-in-mot...
Silviu Din - https://www.pexels.com/@silviu-din-16...
@svetjekolem - https://www.pexels.com/@svetjekolem
José Alfredo Munguía Lira - https://www.pexels.com/@rectorretro
Yaroslav Shuraev - https://www.pexels.com/@yaroslav-shuraev
Danil Shostak - https://www.pexels.com/@danil-shostak...
KATRIN BOLOVTSOVA - https://www.pexels.com/@ekaterina-bol...
Soumya - https://www.pexels.com/@soumya-1446957
Mikhail Nilov - https://www.pexels.com/@mikhail-nilov
Trippy Lagoon - https://www.pexels.com/@trippy-lagoon...
cottonbro studio - https://www.pexels.com/@cottonbro
Oleg Gamulinskii - https://www.pexels.com/@oleg-gamulins...
Engin Akyurt - https://www.pexels.com/@enginakyurt
Kelly - https://www.pexels.com/@kelly
Stas Knop - https://www.pexels.com/@stasknop
Anthony 🙂 - https://www.pexels.com/@inspiredimages
Bedrijfsfilmspecialist.nl - https://www.pexels.com/@bedrijfsfilms...
Dan Cristian Pădureț - https://www.pexels.com/@paduret
Kindel Media - https://www.pexels.com/@kindelmedia
crazy motions - https://www.pexels.com/@crazy-motions...