A Meta-Reinforcement Learning Algorithm for Causal Discovery
What this work does
Trains a recurrent policy across synthetic SCMs to choose budgeted interventions and edit an explicit graph estimate, rewarded by terminal graph accuracy and an intervention bonus.
Connection to discovery
Directly demonstrates that a discovery training problem can learn experiment selection and belief/graph updating jointly, beyond learning a representation alone.
Evidence
Inspected the distinction between intervention and structure actions, known training-graph supervision and transfer to withheld graph structures.
Scope and limitations
The demonstration uses toy acyclic SCMs, a fixed intervention value and privileged graph truth during training. It does not establish broad LLM discovery ability or unrestricted variable invention.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.