CausalARC: Abstract Reasoning with Causal World Models
What this work does
Manually specified SCMs generate ARC-like observational, interventional and counterfactual demonstrations, enabling several reasoning and adaptation protocols.
Connection to discovery
A controlled comparison of reasoning with different causal evidence types and a task-family generalization test.
Evidence
The body separates static few-shot TTT, counterfactual prediction, program synthesis and logical causal-discovery evaluations.
Scope and limitations
Receiving intervention demonstrations differs from actively choosing interventions. Random instances do not by themselves guarantee a novel mechanism family or elimination of structural prior knowledge.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.