From Traces to Agentic Worlds: Agentic Language World Models for Interactive Environment Simulation
Abstract
Realistic environment replicas are increasingly valuable for training and evaluating LLM agents, yet the original systems may be inaccessible or impractical to reproduce. We explore agentic language world modeling: rather than rebuilding an executable environment, a world model agent serves as the environment for a task agent and supports faithful and stateful simulation. We instantiate this paradigm with Trace2Env, a learning-free framework for settings where the original system is unavailable but historical interaction traces remain accessible. Trace2Env reconstructs these traces into a reusable environment worldbook containing environment schemas, grounded evidence, and induced behavioral knowledge. At runtime, the world model agent actively consults the worldbook together with persistent episodic state to infer each action's observation and lasting state effects. Across nine environments, Trace2Env improves both next-observation fidelity and long-horizon interaction consistency over conventional prompt-based LWMs. In multi-turn interaction, task agent actions generated against Trace2Env remain valid more often when replayed in the real environment, indicating that its simulated dynamics better preserve the consequences of earlier actions across successive turns. These results establish agentic language world modeling as an alternative direction for building realistic environment replicas without reconstructing the original executable system.
Community
What if the environment itself were an agent?
- Trace2Env is a training-free framework for agentic language world modeling. It builds world model agents that serve as environments for other task agents, aiming to support faithful, stateful, and long-horizon simulation of textual environments.
- Trace2Env does not recover the original executable system. Instead, it reconstructs environment dynamics from historical action–observation trajectories. During offline reconstruction, these traces are organized into a worldbook containing environment schemas, transition rules, constraints, invariants, grounded evidence, and demonstrations. Check out what a worldbook looks like here.
- At runtime, the world model agent actively calls tools to inspect the worldbook together with the maintained episode state and episodic memory. It infers the next observation together with the state changes that should persist, and a shared runtime harness verifies the proposal and commits the accepted changes, this allows the consequences of earlier actions to carry into later turns.
💡 Such reconstructed environments could be used to train agents without the real system, test agents in safe and reproducible sandboxes, build stateful mocks for tools and APIs, or reconstruct unavailable, private, or legacy environments.
💡 Trace2Env is not tied to a particular environment domain. The paper evaluates it on terminals, software repositories, Android and web apps, enterprise services, and text games. The same framework can be applied to other text-interactive systems, including internal enterprise tools, ticketing workflows, cloud and DevOps consoles, API and MCP backends, and other textual worlds.
🕹️ Try a “fake” Terminal built by Trace2Env:
https://huggingface.co/spaces/Quanyu001/trace2env_demo
⭐ Code: https://github.com/ruyue0001/Trace2Env
🌐 Project: https://quanyulong.net/trace2env/
Get this paper in your agent:
hf papers read 2610.06100 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper