I think open-ended simulation for agents will be a key component for training and planning. Similar as human dreams simulate different scenarios in our head. Biggest challenge will be simulating more abstract and complex systems.
Few months ago I did experiment with an open-ended world simulation for AI agent, where the simulated world was progressively building itself based on each of agent actions in open-ended manner. The idea was to give an agent infinite possibility regarding tool calling, where the tool call would be approved by the adjudicator, and the world state would change. The key issues with the PoC were:
- World decoherence (tried to solve that with a poor graph implementation)
- World flatness - high abstraction did not account for small events that would compound in real world
- Start with empty context was real issue to get the agent to explore the world
Anyways the project came to be really funny when you watched agent struggling in desperation to perform real world actions which would be impossible in real world. Main observation was that when presented agent with current action budget, it modulated the creativity and how desperate its actions were.
I agree; after running out of data on the internet, and humans being too slow to generate data, simulation is the only frontier left for improving things (training, datasets, reasoning). And it's probably the most ethical one too.
If nothing else I'm glad to see "world models" that are actually modeling some kind of worlds, instead of the term being applied as a hype layer for video/splats diffusion.
Give it a day or two and the 'unsloth' people will probably publish a Q6 and Q8 (maybe Q8XL?) quantization in GGUF format for llama-server and other users.
This might be pretty big. One of my biggest frustrations with smaller models (especially MoE) is their failure to track workflow state at a high level. I'm constantly reminding them what we decided on or asking them to revisit, and reminding them eats context.
Seems like this might make that a lot less painful. And if not off the bat, with some minimal tuning or even just good prompting.
I'm a fan of this direction. For me the most interesting use case for these world models isn't even training, it's verification. If this thing or some idealized version of it can actually reliably simulate state transitions, could you use it to verify an agent's execution path against hard constraints and replace/eclipse LLMs-as-a-judge?
Well if you can do this then you don't delegate execution path derivation to the agent. The benefit is a predictable coherent world state where you understand the impact of { current state } x { action } without having to enumerate that huge cartesian product.
A regular LLM acts as a "policy," mapping a current state to a specific action (states → actions). Their new LLM acts as a "world model," mapping a current state and a chosen action to a predicted future state ((states, actions) → subsequent states). Instead of deciding "what to do," its explicit objective is to predict the exact environment observation that will result from the interaction history and the agent's current action.
I assumed at first that it was trained on synthetic data, but they actually went and deployed real physical hosts and virtual machines (e.g. Ubuntu, macOS, and Android) and browsers. They ran agentic systems on these continuously and recorded the actual, real-world interactions.
So it's an LLM that infers next state, or outcome,as structured data e.g. literal HTML code, UI view hierarchies, or accessibility trees.
So, if I'm reading this correctly, whereas a regular LLM would, given a prompt to edit a file, infer a sed call, this "world" model infers the resulting contents of the file.
Here's the description of the world model prompt for the web domain: "A precise GUI state simulator — given the current screen (as HTML) and a user action, predicts the exact next screen as a complete, self-contained HTML document." (You can click the world model prompt box to expand it and see the full prompt.)
So the world model generates the current state (an html document), an agent tells it what action it wants to perform, the world model generates the next state (another html document).
The other domains are similar, but w/ domain-specific nuance.
Same thing, but qwen has decided to rebrand certain LLMs that were trained slightly differently as "world models". Despite the fact that "world model" typically means !LLM.
I believe the benchmark listed is about simulating the environment for the various tasks, rather than doing them. It seems that the point of this model is to generate sim data to improve other models with
> Figure 1: Overview of Qwen-AgentWorld. Top: Qwen-AgentWorld is a unified native language world
model across seven domains. Bottom: We explore two complementary strategies for applying world
modeling to enhance language agents (mainly using the 35B-A3B model as agent): Decouple and Unify ,
where the world model serves as the environment simulator and agent foundation model, respectively.
The bars above the label "Infinite Real-World Envs" show growth for example from approx 42 to 55 but the red label says "+7.1". It's wrong for all of them.
Few months ago I did experiment with an open-ended world simulation for AI agent, where the simulated world was progressively building itself based on each of agent actions in open-ended manner. The idea was to give an agent infinite possibility regarding tool calling, where the tool call would be approved by the adjudicator, and the world state would change. The key issues with the PoC were:
Anyways the project came to be really funny when you watched agent struggling in desperation to perform real world actions which would be impossible in real world. Main observation was that when presented agent with current action budget, it modulated the creativity and how desperate its actions were.If nothing else I'm glad to see "world models" that are actually modeling some kind of worlds, instead of the term being applied as a hype layer for video/splats diffusion.
https://huggingface.co/Qwen/Qwen-AgentWorld-35B-A3B
Seems like this might make that a lot less painful. And if not off the bat, with some minimal tuning or even just good prompting.
I assumed at first that it was trained on synthetic data, but they actually went and deployed real physical hosts and virtual machines (e.g. Ubuntu, macOS, and Android) and browsers. They ran agentic systems on these continuously and recorded the actual, real-world interactions.
So it's an LLM that infers next state, or outcome,as structured data e.g. literal HTML code, UI view hierarchies, or accessibility trees.
Here's the description of the world model prompt for the web domain: "A precise GUI state simulator — given the current screen (as HTML) and a user action, predicts the exact next screen as a complete, self-contained HTML document." (You can click the world model prompt box to expand it and see the full prompt.)
So the world model generates the current state (an html document), an agent tells it what action it wants to perform, the world model generates the next state (another html document).
The other domains are similar, but w/ domain-specific nuance.
> Figure 1: Overview of Qwen-AgentWorld. Top: Qwen-AgentWorld is a unified native language world model across seven domains. Bottom: We explore two complementary strategies for applying world modeling to enhance language agents (mainly using the 35B-A3B model as agent): Decouple and Unify , where the world model serves as the environment simulator and agent foundation model, respectively.
Where is the mistake?
The bars above the label "Infinite Real-World Envs" show growth for example from approx 42 to 55 but the red label says "+7.1". It's wrong for all of them.
(For another example, the charts in the August 2025 GPT-5 presentation)
https://github.com/QwenLM/Qwen-AgentWorld
https://huggingface.co/Qwen/Qwen-AgentWorld-35B-A3B