Causal Representation Learning

We propose a causal representation learning framework to understand the hierarchical structure of language model capabilities, revealing causal relationships between general problem-solving, instruction-following, and mathematical reasoning abilities.