From Traces to Agentic Worlds: Agentic Language World Models for Interactive Environment Simulation

From Traces to Agentic Worlds. Agentic Language World Models for Interactive Environment Simulation. By Quanyu Long and six others, from Nanyang Technological University and the Hong Kong Polytechnic University. Published on the fifth of October. Its code repository is worked out: about one hundred and seventy source files, with a readme and a licence.

Realistic replicas of software environments are increasingly valuable for training and evaluating language model agents. The trouble is that the original systems may be inaccessible, or impractical to reproduce. Sometimes all that survives of a system is a record of how people and agents interacted with it: historical traces, with no executable environment left to run.

The authors explore what they call agentic language world modeling. Rather than rebuilding an executable environment, a world model agent serves as the environment for a task agent, and supports simulation that is both faithful and stateful. They instantiate the idea as Trace2Env, a learning-free framework aimed squarely at the case where the original system is unavailable but its interaction traces remain.

Trace2Env works in two phases. First it reconstructs the traces into a reusable environment worldbook. The worldbook holds environment schemas, grounded evidence, and induced behavioral knowledge about how the environment tends to respond. Then, at runtime, the world model agent actively consults that worldbook together with a persistent episodic state. From both, it infers what observation each action should produce and what lasting effect the action has on the state. The episodic state is what lets one turn's consequences carry into the next.

Across nine environments, Trace2Env improves both next-observation fidelity and long-horizon interaction consistency over conventional prompt-based language world models. The multi-turn test is the telling one. Actions a task agent generated against Trace2Env remain valid more often when they are replayed in the real environment. The authors read this as evidence that the simulated dynamics better preserve the consequences of earlier actions across successive turns.

Here a replica is something an agent maintains, not a program someone rebuilds. Give a world model a structured memory of past traces and a running state, and it can stand in for a system that no longer runs. The authors present this as an alternative direction for building realistic environment replicas without reconstructing the original executable system.