Pre-print proposes causal world models to improve modular LLM agents in structured tool environments

LLM agents increasingly operate through modular systems—such as order, payment, inventory, and shipment services—where actions in one module alter valid transitions in another. Standard world models trained on observational traces struggle to support intervention-time planning because observational data may show that payment precedes shipment without revealing whether payment authorizes shipment, inventory mediates the effect, or a hidden trigger explains both.

To address this, the authors propose FedCausalCompose, a causal world-model framework for modular LLM agents that uses local actions to collect intervention-response evidence for cross-module interfaces. They demonstrate that observational world models incur an irreducible interventional error when unblocked back-door paths are present, and that interface recovery improves as intervention-response coverage increases.

Under controlled conditions, an oracle causal composition can outperform the non-causal lower bound when coverage and local mechanism errors are constrained. The authors then test these predictions in diagnostic agent settings, finding that causal interfaces provide the most benefit in structured tool environments where API signatures expose preconditions and downstream effects.

In contrast, dialogue and narrative environments often fail to leverage raw edge lists unless a short attention anchor makes the causal information decision-relevant. The results highlight a practical condition: causal structure helps modular LLM agents when cross-module interfaces are both statistically identifiable and presented in a form the agent can use at action time.