Agentic World Modeling for Agent Test-time Scaling without Environment Replica

| Source: arXiv AI

Tags: Agentic World Modeling, Test-Time Scaling, tau-bench, AppWorld, OfficeBench, LLM Agents

Microsoft-affiliated researchers propose Agentic World Modeling, which lets LLM agents run test-time scaling over one shared, non-resettable environment by simulating actions through read-only queries, beating a single-shot simulation baseline by 40% on average.

Details

Test-time scaling for agents usually assumes candidate trajectories can be executed in parallel, which breaks in shared enterprise workspaces where actions persist and environments cannot be cloned. Agentic World Modeling (AWM) turns next-observation prediction into a multi-turn evidence-gathering task: a simulator queries the real environment through safe read-only interfaces and computes the resulting observation without changing shared state. Across tau-bench, AppWorld and OfficeBench, AWM beats a single-shot simulation baseline by 40.07% on average and a trained world-model baseline by 12.36%. For teams deploying agents against production systems, it points to a way to explore options without side effects.