AI News

The paper studies whether causal world models improve planning for LLM agents that operate across modular services such as ordering, payment, inventory, and shipment. Observational traces can reveal that events occur in sequence, but cannot establish whether one action enables another, whether an intermediate module mediates the effect, or whether a hidden trigger explains both. To address this gap, the authors introduce FedCausalCompose, which uses local action-intervention responses to recover causal interfaces between modules. They show that observational world models retain irreducible interventional error when back-door paths remain unblocked, while interface recovery improves with broader intervention-response coverage. Under controlled coverage and local-mechanism errors, an oracle causal composition can outperform the non-causal lower bound. Diagnostic agent experiments find the largest benefits in structured tool environments, where API signatures expose preconditions and downstream effects. Dialogue and narrative environments often fail to use raw causal edge lists, but a short attention anchor can make the information relevant to decisions. The study concludes that causal structure helps when cross-module interfaces are statistically identifiable and presented in a form the agent can use at action time.