Researchers at Nanyang Technological University Singapore introduced Trace2Env, a training-free framework that reconstructs interactive text environments from historical action and observation traces. The paper, published on October 5 and submitted to Hugging Face Papers on October 9, proposes using an agent as the simulated environment when the original system is inaccessible or impractical to reproduce.

Trace2Env organizes past interactions into a “worldbook” containing environment schemas, transition rules, constraints, invariants, evidence and demonstrations. During use, a world model agent consults that record alongside persistent episode state and memory to infer what the task agent observes next and which changes should persist. A shared runtime harness checks the proposed changes and commits accepted updates, carrying earlier actions’ consequences into later turns.

The researchers evaluated Trace2Env across nine environments, including terminals, software repositories, Android and web apps, enterprise services and text games. They report improved next-observation fidelity and long-horizon interaction consistency over conventional prompt-based language world models. In multi-turn tests, actions generated against Trace2Env remained valid more often when replayed in the real environment, which the authors say indicates that its simulated dynamics better preserve the effects of earlier actions across turns.

The authors say reconstructed environments can support agent training without access to the real system, testing in safe and reproducible sandboxes, stateful mocks for tools and APIs, and reconstruction of unavailable, private or legacy environments. They say the framework is not tied to one domain and can apply to other text-interactive systems, including ticketing workflows, cloud and DevOps consoles, and API and MCP backends. The paper presents agentic language world modeling as an alternative for building environment replicas without rebuilding the original executable system.

Researchers at Nanyang Technological University Singapore introduced Trace2Env, a training-free framework that reconstructs interactive text environments from historical action and observation traces. The paper, published on October 5 and submitted to Hugging Face Papers on October 9, proposes using an agent as the simulated environment when the original system is inaccessible or impractical to reproduce.

Trace2Env organizes past interactions into a “worldbook” containing environment schemas, transition rules, constraints, invariants, evidence and demonstrations. During use, a world model agent consults that record alongside persistent episode state and memory to infer what the task agent observes next and which changes should persist. A shared runtime harness checks the proposed changes and commits accepted updates, carrying earlier actions’ consequences into later turns.

The researchers evaluated Trace2Env across nine environments, including terminals, software repositories, Android and web apps, enterprise services and text games. They report improved next-observation fidelity and long-horizon interaction consistency over conventional prompt-based language world models. In multi-turn tests, actions generated against Trace2Env remained valid more often when replayed in the real environment, which the authors say indicates that its simulated dynamics better preserve the effects of earlier actions across turns.

The authors say reconstructed environments can support agent training without access to the real system, testing in safe and reproducible sandboxes, stateful mocks for tools and APIs, and reconstruction of unavailable, private or legacy environments. They say the framework is not tied to one domain and can apply to other text-interactive systems, including ticketing workflows, cloud and DevOps consoles, and API and MCP backends. The paper presents agentic language world modeling as an alternative for building environment replicas without rebuilding the original executable system.