JING and DAO want to simulate a world that continues to exist off-screen

XGEN Labs combines JING, an interactive first-person model, with DAO, an engine responsible for maintaining the state of a shared environment. The prototype seeks to move beyond frame-by-frame video generation to preserve an evolving world among multiple agents.

A world that does not stop with the camera

Looking around, walking through a room, or starting a conversation is only one part of the system. The simulation prototype developed by XGEN Labs separates the task between two components: JING generates what an agent sees and hears from its own perspective, while DAO maintains the broader state of the environment, its rules, and the consequences of actions taking place inside it.

That separation is meant to let the simulated world keep changing beyond what a user can directly observe. DAO can track events occurring elsewhere in the environment and then provide JING only with information that should be perceptible from the current observer's position. Multiple agents can therefore operate within the same space without continuously accessing its complete state. JING turns actions into video and audio

The publicly available component is XGEN-JING, an egocentric interactive experience model built on MiniMax-H3. Given actions, reference images, and observation history, it jointly generates video and audio for navigation, object interaction, and conversation.

Camera movement can be controlled from the keyboard, while text instructions guide selected interactions and dialogue. Reference images can define characters, objects, and environments before exploration begins, allowing the same initial scene to develop differently depending on subsequent actions.

The JING-Flash-v1 version available in the project's GitHub repository uses four-step bidirectional inference. XGEN also provides inference code, examples, and Prompt skills that convert a story and reference images into structured test cases. A causal version of the model and a full technical report are still listed as upcoming.

Local inference remains demanding. The demonstration setup was validated on six H100 GPUs, split between the text encoder, video and audio components, and four GPUs running the DiT with sequence parallelism. An early result on WBench

On WBench, a benchmark for interactive video World Models, the bidirectional version of XGEN-JING recorded an average score of 81.0 on the 289-case Full split when it entered the leaderboard, placing it first at the time. On the 158-case navigation split, its 81.9 score placed it second.

The leaderboard has continued to change since then, so those positions are better treated as a snapshot of the benchmark at the time of evaluation rather than a permanent ranking. WBench evaluates 22 metrics spanning video quality, setting adherence, interaction, consistency, and physical behavior, and explicitly notes that no evaluated model dominates every dimension. From generating a scene to maintaining a world

The more ambitious part of the prototype sits on the DAO side. JING produces an agent's perceptible experience, while DAO is intended to preserve what exists behind it: shared state, rules, environmental changes, and decisions made independently by other agents.

XGEN connects this architecture to its broader OASIS project, envisioned as a simulation where a world, and potentially a society of agents, can continue to evolve with or without a human observer. For now, the public release mainly exposes JING. Detailed technical documentation for DAO and the complete coupled system has not yet been released.