CausaLab: A Scalable Environment for Interactive Causal Discovery Toward AI Scientists
What this work does
Samples hidden SCMs, permits bounded interventions and asks agents for both held-out predictions and an explicit graph/mechanism description.
Connection to discovery
A close benchmark for measuring whether successful answers are backed by recovered mechanisms and appropriate experiments.
Evidence
Comparisons distinguish task accuracy from edge and mechanism fidelity, including cases where topology is preserved but changed functional forms break prediction.
Scope and limitations
The variables, mechanisms and intervention interface are generated within a known family. High prediction accuracy is insufficient evidence of mechanism recovery; real-world causal sufficiency remains outside the simulator.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.