Agentic World Modeling for Agent Test-time Scaling without Environment Replica
Wendi Li, Sharon Li and colleagues (University of Wisconsin–Madison and Microsoft) propose Agentic World Modeling (AWM), which lets agents run test-time scaling in shared environments that cannot be copied or reset by simulating each candidate action with read-only queries to the real environment.
Ask this paper
Problem. Trajectory-level test-time scaling assumes candidate trajectories can be executed in parallel. Enterprise workspaces and cloud services are shared and persistent, and actions can have irreversible side effects.
Method. AWM treats next-observation prediction as a multi-turn evidence-gathering task. For each hypothetical action, a simulator agent queries the real environment through safe read-only interfaces, reasons over the results and computes the resulting observation without changing shared state.
Results. Across tau-bench, AppWorld and OfficeBench, AWM outperforms a single-shot simulation baseline by 40.07% on average and a trained world-model baseline by 12.36% on average.
Use. Combined with trajectory-level test-time scaling, AWM allows parallel exploration over a single shared environment.
Abstract
Test-time scaling (TTS) improves large language model agents by sampling and selecting among multiple candidate trajectories, but common realizations assume that these trajectories can be executed in parallel. This assumption breaks in realistic deployments, where enterprise workspaces and cloud services are shared and persistent, costly or infeasible to replicate, and actions may introduce irreversible side effects or cross-trajectory interference. Existing generative world models avoid direct execution, but typically compress environment state into a single prompt, limiting their ability to reliably simulate transitions under partial observability. Instead, we introduce Agentic World Modeling (AWM), which reformulates next-observation prediction as a multi-turn evidence-acquisition task. For each hypothetical action, an agentic simulator adaptively queries the real environment through safe read-only interfaces, reasons over the returned evidence, and uses executable computation to construct the resulting observation without mutating shared state. Integrated with trajectory-level TTS, AWM enables parallel exploration over a single shared environment. Across $τ$-bench, AppWorld, and OfficeBench, AWM outperforms a single-shot simulation baseline by 40.07% on average, and surpasses a trained world-model baseline by 12.36% on average. These results establish interactive, read-grounded simulation as a practical foundation for safe and scalable agentic test-time scaling in shared, non-resettable environments.