When Should LLM Semantic Priors Enter Partially-Identified Causal Discovery?

Summary:This study separates the use of LLM semantic priors in partially identified causal discovery into three decisions: where to allocate semantic queries, which directional proposals to admit, and which candidate DAGs to retain. It distinguishes population-level non-identifiability from finite-sample instability and compares semantic orientations with statistical and matched-random controls. Across benchmark networks, semantic information improves DAG recovery over coin-flip orientation, but priori

---

Research Question: Useful Semantic Priors Are Not Automatically Reliable Causal Decisions

With observational data alone, multiple directed acyclic graphs (DAGs) can share the same skeleton and conditional-independence structure. Consequently, some edge orientations are population-level non-identifiable, while finite samples can make otherwise recoverable orientations unstable. LLMs can propose directions using variable names, domain descriptions, and common sense, but an incorrect proposal may exclude the true DAG before downstream inference begins.

The study's main contribution is to decompose LLM integration into three independently testable decisions:

1. Query allocation: Which edges should receive a limited number of semantic queries?
2. Direction admission: Which proposed orientations have enough evidence to become hard constraints?
3. DAG retention: How should the pipeline tolerate some incorrect directions while preserving graphs that may contain the truth?

This structure is more engineering-friendly than asking an LLM to emit an entire causal graph, because each stage can be monitored, replaced, and rolled back independently.

Mechanism: Separating Population Non-Identifiability from Finite-Sample Instability

When two observational distributions satisfy the same conditional-independence relations, changing some edge directions need not change the distribution. This is population non-identifiability. Prioritizing unstable edges for semantic queries implicitly assumes that frequently flipped edges are also those that semantic knowledge can correct most effectively.

That assumption need not hold. Instability may reflect finite-sample noise rather than a fundamental ambiguity, and a language model may repeatedly propose the same incorrect direction on the same class of edges. The experiments therefore compare semantic orientation with statistical orientation, random orientation, and matched-random query allocation.

For a fair comparison, the LLM query budget, prompt template, and response parser must remain fixed. Otherwise, an apparent gain from targeted queries could simply reflect additional inference calls.

The admission mechanism can be framed as a calibrated selection problem. For a candidate orientation g, the pipeline computes a statistic or certificate T(g) and chooses a threshold τ, admitting only directions satisfying T(g) ≥ τ. Ideally, τ controls the false-admission rate and bounds the probability of excluding the true DAG. However, reliable calibration requires coverage of incorrect orientations, boundary cases, and near-equivalent graph structures. The study finds that the available calibration tasks are insufficient to support a nontrivial admission threshold. LLM confidence therefore cannot be treated as a certified probability.

Retention: Candidate Graphs Are Not Automatically Useful

DAG retention must balance recall against precision. A strict threshold can remove the true graph, while a loose threshold propagates many wrong orientations into the candidate set. Error-tolerant retention mechanisms are therefore useful, but their evaluation depends on a valid reference collection.

This reference requirement is a major bottleneck. If a learned reference collection does not contain the true DAG, a large retained candidate set does not prove that the truth was preserved. In the reported experiments, learned reference sets contain the true DAG in only three of 80 planned tasks, substantially weakening downstream interpretation.

Findings: Semantics Helps, but Targeting Does Not

Across benchmark networks, semantic orientation improves DAG recovery over coin-flip orientation. This indicates that variable semantics contribute information unavailable to an uninformative random baseline. However, directing queries toward unstable edges offers no established advantage over matched-random allocation. Statistical instability is therefore not a reliable proxy for where semantic queries are most valuable.

The admission results are more cautious. The tested certificate cannot support meaningful nontrivial admission with the available calibration tasks. A system can record LLM proposals, but the reported evidence does not establish a reliable statistical guarantee on the risk of excluding the true graph.

The reference-set result is another important negative finding. With the true DAG present in only three of 80 planned tasks, the current learned graph references have limited coverage. Downstream retention results therefore should not be interpreted as proof of reliable causal discovery on unseen data.

This does not mean that LLMs are useless for causal discovery. The narrower conclusion is that semantic information is better than an uninformative random baseline, but the experiments do not yet support an efficient, calibrated, and provably safe automatic integration pipeline.

Engineering Trade-offs and Reproducibility Risks

Semantic querying incurs prompt-inference, entity-to-variable alignment, response parsing, caching, and repeated-call costs. Querying every candidate edge scales as O(|E|) in the number of edges. Whole-graph prompting reduces the number of calls but creates larger contexts, allows proposed directions to interact, and makes edge-level auditing harder.

The matched-random control is consequently useful in practice. When targeted queries show no stable gain, random or stratified-random allocation is cheaper and avoids relying on unstable-edge estimates as if they were measures of semantic value.

Evaluation should also control for:

• Benchmark-graph bias: Synthetic networks may use names and descriptions that favor models trained on similar language;
• Prompt sensitivity: Anonymization, synonyms, or adversarial descriptions can change proposed directions;
• Model-version bias: Different LLMs have different commonsense coverage and prompt sensitivity, shifting calibrated thresholds;
• Selection bias: Reference DAGs that are easy to describe can exaggerate the apparent value of semantic priors.

The paper does not establish end-to-end reliability on large real-world causal systems. In drug discovery, public policy, or industrial troubleshooting, beating a random baseline alone is insufficient for autonomous intervention recommendations.

Practical Deployment Guidance

The defensible near-term role of an LLM is that of a proposal generator, not an irreversible causal adjudicator:

1. Use statistical methods to produce the skeleton, candidate orientations, and stability estimates.
2. Record each LLM proposal with its prompt, model version, raw response, and parsed result.
3. Estimate false-admission rates on an independent calibration set rather than treating LLM confidence as probability calibration.
4. Maintain error-tolerant DAG candidate sets and continuously measure true-graph coverage.
5. If targeted querying does not consistently beat matched-random allocation, use random or stratified-random budgets instead of interpreting instability as a semantic-value signal.
6. Keep human review and intervention-data feedback loops for high-stakes decisions.

The study's practical contribution is a clear boundary: LLM semantic priors can supply additional information, but query allocation, direction admission, and graph retention require independent calibration and coverage evidence. A language model's commonsense cannot simply be promoted into causal fact.

---

Source: ArXiv AI (cs.AI) (2026-11-13)

— END —

Tags: #LLM Semantic Priors Follow #Causal Discovery Follow #DAG Follow #Calibration Follow #Uncertainty Follow