Learning to Induce Causal Structure
What this work does
A transformer maps observational/interventional datasets to an autoregressive distribution over graph adjacency entries, trained on synthetic graph/data pairs.
Connection to discovery
A learned inference baseline that separates recovering structure from selecting which intervention data to acquire.
Evidence
Inspected the learned dataset-to-graph mapping, simulator assumptions, lack of an explicit acyclicity constraint and synthetic/generalization comparisons.
Scope and limitations
Finite-data graph predictions inherit the training distribution and may contain cycles. Supervised synthetic graph recovery does not establish unrestricted causal discovery.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.