AI

When Should Semantic Priors Enter Partially-Identified Causal Discovery?

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 13/11/2026, 07:48:24 UTC.

Briefings

A new framework analyzes LLMs' ability to resolve causal directions in observational data, distinguishing between population and sample instability. Experiments show semantic information improves recovery over random guesses but targeting unstable edges offers no advantage. The current admission certificate lacks calibration for nontrivial admission, and downstream retention depends heavily on reference adequacy, where only 3 of 80 planned tasks recovered the true DAG.

Key points

• Semantic information improves causal direction recovery in observational data compared to random guesses.
• Targeting unstable edges provides no advantage over matched-random allocation for causal discovery.

Context

This research addresses a critical gap in causal discovery where LLMs can provide semantic information to resolve ambiguous causal directions, yet incorrect proposals can invalidate downstream inference. The study evaluates three key decisions: query allocation, proposal admission, and graph retention.

Limitations

The tested admission certificate cannot support nontrivial admission with available calibration tasks. Downstream retention depends on reference adequacy, where only 3 of 80 planned tasks recovered the true DAG, highlighting significant uncertainty in the current framework's reliability.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 13/11/2026, 07:58:46 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeigkgnlx77xotmqnjqjezmfoa2illylt7ect37l2fhc55gm2sizbi4