Ask the World Before Acting: Environment Probing for Calibrating Agent World Models
Long-horizon language agents often fail because their internal world models become outdated, leading them to act on incorrect assumptions about the environment. This work introduces EnvProbe, a framework where agents actively query the real environment to verify and correct inaccurate beliefs during execution. By scoring belief fields based on task relevance and staleness, EnvProbe improves terminal accuracy significantly over methods that rely only on internal reasoning.