Ask the World Before Acting: Budgeted Environment Probing for World-Model Calibration

2026-06-30

AI Summary

This work addresses the challenge of belief drift in long-horizon language agents, which often fail due to gradual misalignment between their internal world models and the actual environment. To mitigate this issue, the authors propose a budget-constrained environmental probing mechanism that treats interactions with the environment as a scarce calibration resource. Before executing actions, the agent actively queries critical belief fields and updates its model accordingly. The approach introduces a novel type-hierarchical probing strategy that differentiates calibration methods for procedural and spatial beliefs, guided by a structured belief table and task structure to inform probe selection. Experimental results demonstrate that incorporating environmental evidence at intermediate planning stages significantly reduces terminal world model error, effectively alleviating belief drift in extended tasks.

Abstract

Long-horizon language agents do not only choose actions; they carry a private model of the world from one decision to the next. When that model drifts, a later failure can be decided before the failing action is ever taken. We study a direct repair mechanism: before committing to the next task action, an agent may ask the environment about one belief field and write the answer back into its world model. This makes environment interaction a scarce calibration resource, not merely a way to advance the task. We introduce \method, a budgeted probing operator for structured belief tables. The useful probes are not the same everywhere. Procedural beliefs, such as tool dependencies, can often be repaired by targeted checks, but those checks spend steps that the task may need. Spatial beliefs, such as object locations and graph edges, rely more on structural cues; the agent's own confidence can be a poor guide when the world changes off-screen. A type-stratified analysis formalizes this probe-action frontier, and controlled experiments show that mid-planning environment evidence reduces terminal world-model error when the probe policy follows the structure of the task.

Problem

• world-model calibration
• budgeted probing
• language agents
• belief repair
• environment interaction

Innovation

• budgeted probing
• world-model calibration
• structured belief tables
• type-stratified analysis
• language agents