LLM agents are increasingly built on modular systems such as order, payment, inventory, and shipment services. An action in one module changes which state transitions are valid in another.
Standard world models only fit observational traces, which is insufficient for intervention‑time planning. A trace may show that payment precedes shipment, but it does not reveal whether payment authorizes shipment, inventory mediates the effect, or a hidden trigger explains both.
We introduce FedCausalCompose, a causal world‑model framework that uses intervention‑response evidence from local actions to recover cross‑module interfaces.
Our analysis shows that observational world models incur an irreducible interventional error when unblocked back‑door paths exist; interface recovery improves with greater intervention‑response coverage; and, under controlled coverage and local mechanism error, an oracle causal composition can surpass the non‑causal lower bound.
We validate these predictions in diagnostic agent settings. Causal interfaces provide the most benefit in structured tool environments, where API signatures expose preconditions and downstream effects. In dialogue and narrative environments, raw edge‑list information is often ignored unless a short attention anchor makes the causal information decision‑relevant.
The takeaway is that causal structure helps LLM agents when cross‑module interfaces are both statistically identifiable and presented in a form the agent can use at action time.
Review