2026 Beijing BAAI Conference Opens: BAAI Releases World Models, Defining AI's Paradigm Shift from 'Predicting the Next Token' to 'Predicting the Next Physical State'

On June 12, 2026, the 8th Beijing BAAI Conference opened at Zhongguancun International Innovation Center. The conference features two Turing Award winners, over 40 AI company CEOs and chief scientists, more than 200 top experts and scholars, and over 30 young scientists under 30. It marks the first time that China's most representative innovations in world models and agents are showcased together.

Key Releases: World Models and Paradigm Shift

BAAI President Wang Zhongyuan presented the 2026 progress report, systematically describing AI's paradigm shift from 'predicting the next token' to 'predicting the next physical state.' BAAI was the first institution in China to propose and research world models. In 2023, Yann LeCun elaborated on the concept at BAAI Conference, and in 2024, BAAI identified world models as the next-generation large model technology.

Two world models were released:

Wujie·Physis-v0.1: The world's first general-purpose world foundation model, using physical latent space representation instead of traditional pixel or frame-level prediction. It supports long-range reasoning in over 50 complex physical scenarios, with four core capabilities: physical consistency, action causality, long-range predictability, and general generalization. Applications include serious industry, embodied intelligence, physical simulation, and scientific research.
Wujie·RoboBrain Orca (under development): An embodied brain centered on predicting the next physical state, integrating 'thinking, seeing, and moving' to simultaneously generate language reasoning, visual prediction, and action decisions, supporting long-term autonomous operations in real environments like logistics and hotel services.

BAAI categorizes current world model approaches into four types: language-centric (e.g., VLM/VLA), pixel-centric (e.g., Sora), 3D structure-centric (e.g., World Labs Marble), and visual representation-centric (e.g., JEPA). BAAI believes that true world models should focus on 'predicting the next physical state,' covering all modalities and possessing active interaction capabilities.

Other Important Achievements

Wujie·Emu3: Published in Nature in January 2026, the first large model result led by a Chinese research institution to appear in Nature. Emu3.5 completed key upgrades, extending from 'predicting the next token' to 'next state prediction' across visual and language sequences.
Wujie·Brainμ1.0: The world's first multimodal neuroscience large model unifying understanding and generation, encoding cross-species, full-modality brain signals into standard tokens. Its research on 'memory-sleep' regulation mechanisms was published in Science.
Wujie·OpenComplex2.5: A generalizable, physically realistic next-generation AI-driven drug discovery model, unifying four key steps: pocket identification, reverse screening, structure prediction, and affinity prediction.
FlagOS2.1: Supports 32 chips from 18 manufacturers, the most extensive computing system software stack globally, with over 80 ecosystem members and over 375,000 global downloads.

Guest Lineup and Core Topics

2015 Turing Award winner Whitfield Diffie focused on security challenges in the agent era, emphasizing that confinement of agents remains insufficient in current programming. 2024 Turing Award winner Andrew Barto, in his talk 'Rediscovering Reinforcement Learning,' pointed out that computational research on reinforcement learning and neural networks has been intertwined from the start, and warned of the risk of 'abnormal instantiation' in reward signal design.

BAAI Chairman Huang Tiejun and Zhejiang Lab Director Wang Jian engaged in a dialogue titled 'Someone Is Always Ten Years Ahead of Their Time,' discussing the evolution of China's AI paradigm, new paradigms in the agent era, and the relationship between humans and AI. In the roundtable 'Reconstructing the World: China's Large Model Summit Dialogue,' Wang Zhongyuan, Zhu Jun, Luo Fuli, Liu Zhiyuan, and An Bo discussed super model capability evolution, AI self-evolution, multimodality, and world models.

Young Talent and Organizational Structure

The conference announced that Chen Boyuan, a 22-year-old scholar from Peking University and founder of Inverse Matrix Technology (Physis), will lead BAAI's Behavioral World Model Innovation Center. Chen won the ACL 2025 Best Paper during his undergraduate studies, has published multiple papers at top conferences like NeurIPS and ACL, and has over 2,000 Google Scholar citations. His company, Inverse Matrix Technology, focuses on general-purpose world foundation models and secured over $10 million in funding at inception.

Open-Source Ecosystem and Industry Impact

Since its establishment in 2018, BAAI has released over 200 open-source models, with cumulative global downloads exceeding 1 billion, and has incubated numerous startups in large models and embodied intelligence. This conference showcases the technological leap from 'Wudao' to 'Wujie,' building a complete technical path toward physical AGI.