When Do Causal World Models Help Modular LLM Agents
Not provided
Abstract
This paper explores the effectiveness of causal world models in modular LLM agents, particularly in environments where actions in one module affect transitions in another.
Reality Card
Causal world models improve intervention-time planning in modular LLM agents by providing evidence for cross-module interfaces, particularly in structured tool environments.
Causal interfaces significantly enhance performance in structured tool environments, with an oracle causal composition outperforming non-causal lower bounds when local mechanism errors are controlled.
The effectiveness of causal models is limited in dialogue and narrative environments where causal information may not be utilized effectively.
Paper to code
Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.