Do Enterprise Systems Need Learned World Models? The Importance of Context to Infer Dynamics

Researchers question whether enterprise agents still need learned world models when system rules are readable at inference time.

They introduce enterprise discovery agents that inspect live configuration and build predictions from the active system instance.

CascadeBench shows offline-trained world models perform well in-distribution but degrade under deployment shifts.

Discovery-based agents stay more reliable because they ground transition dynamics in current tenant-specific business logic and runtime rules.