When Should Semantic Priors Enter Partially-Identified Causal Discovery?
Summary
- Observational causal discovery often leaves causal directions unresolved or unstable, prompting the use of large language models for semantic priors.
- Researchers study three decisions: where to allocate semantic queries, which proposals to admit, and which causal graphs to retain.
- Experiments show semantic information improves recovery over coin-flip orientations, but targeting unstable edges has no established advantage over matched-random allocation.
- Key limitations identified include calibration size, reference adequacy, and the failure of learned references to contain the true DAG in most tasks.
Why it matters
Understanding the limits and mechanics of integrating semantic priors from LLMs into causal discovery helps researchers avoid flawed downstream inference caused by unreliable allocation and reference adequacy.