Causal Reasoning from Meta-reinforcement Learning
What this work does
Meta-trains a recurrent policy to gather information before a rewarded causal query, with within-episode adaptation carried by recurrent state rather than new weight updates.
Connection to discovery
A concrete training formulation for choosing informative interventions and using their results to answer later causal questions.
Evidence
Synthetic causal networks vary observational, conditional and interventional access, including hidden common causes. Comparisons isolate when associative strategies are insufficient.
Scope and limitations
Graph size, mechanism family, actions and query protocol are designed. Transfer within that generator is narrower than unbounded scientific discovery.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.