From Wudao to Wujie: Zhiyuan Institute Advances AI, Physical‑World, and Life‑Science Integration at the 2026 Beijing Conference

The 8th Beijing Zhiyuan Conference opened on June 12, 2026 at the Zhongguancun International Innovation Center, gathering more than 200 top AI researchers, CEOs, and scholars, including Whitfield Diffie and Andrew Barto, to discuss frontier AI topics such as agents, world models, and embodied intelligence.

In his 2026 research progress report, Institute Director Wang Zhongyuan highlighted breakthroughs in three core areas—base models, embodied agents, and foundational hardware‑software ecosystems—and announced the latest open‑source developments.

Base models : The institute released the multimodal base model Emu 3.5 , which extends next‑token prediction to unified text‑image‑video learning and set new Chinese multimodal records after its Nature publication in January 2026. New models include Brainμ 1.0 , the first unified multimodal neuroscience model that aligns EEG, fMRI, MEG, fNIRS, calcium imaging, and language‑image‑video tokens; the AI‑driven drug‑discovery model OpenComplex 2.5 , which accurately predicts IDP conformations and covers the four key steps of pharmaceutical pipelines; and the world‑model base Physis‑v0.1 , the first general‑purpose model that predicts the next physical state, enabling long‑range, causally consistent reasoning across 50+ complex physical scenarios.

The institute also classified existing world‑model research into four categories—language‑centered, pixel‑centered, 3D‑structure‑centered, and visual‑representation‑centered—explaining why only the fourth category captures true physical dynamics.

Embodied agents : To address hardware immaturity, data scarcity, limited model capability, and deployment difficulty, Zhiyuan built a full‑stack embodied‑intelligence ecosystem, releasing RoboBrain , RoboOS , and the upcoming RoboBrain Orca that predicts the next physical state using massive ego‑centric interaction data. Four domain‑specific agents were demonstrated: BAAI Cardiac Agent (CMR‑assistant with AUC > 0.93), AREX (autonomous scientific‑research agent), SoulAgent (personal assistant that reduces token cost by 30% and resource usage by 80%), and a risk‑discovery agent for harmful proteins that closes the loop between computational simulation and wet‑lab validation.

FlagOS 2.1 : The open‑source FlagOS stack now supports 18 chip vendors and 32 chip models, offering over 600 operators, unified compilers for 18 vendors, and a common communication library for 12 vendors. It enables day‑zero multi‑chip deployment of models such as DeepSeek V4, Qwen 3.6, MiniMaX M2.7, MiniCPM‑o 4.5, and others, and includes performance‑optimizing components like Triton‑TLE language extensions and the FlagCX communication library.

Keynote speeches: Whitfield Diffie presented "Security For AI Agents" and argued for formal methods to improve AI reliability, emphasizing the need for stronger confinement of agents. Andrew Barto revisited reinforcement learning, describing it as a triad of control, search, and memory, and warned about reward‑signal mis‑specification risks.

Panel discussions and podcasts explored AI’s paradigm shift from digital to physical worlds, the emergence of agents as autonomous systems, and the long‑term coexistence of humanity and AI, with insights from Huang Tiejun, Wang Jian, and other leading Chinese AI figures.

In conclusion, Zhiyuan Institute positions itself as a pioneer in world‑model research, open‑source ecosystem building, and embodied‑agent development, aiming to drive the next AI paradigm—physical‑world‑aware artificial general intelligence.