Causal Representation Learning

Causal representation learning aims to recover meaningful variables and relationships from complex observations, so models can predict what happens under interventions and transfer across changing environments. Approaches combine latent-variable models and nonlinear independent component analysis with disentanglement, causal graphs, and invariant learning across environments, interventions, or time. Work spans identifiability theory and methods for images, scientific time series, graphs, and action-conditioned world models. A central tension is that recovery guarantees often rely on assumptions that real data may violate, while predictive accuracy alone does not ensure that a representation captures causal structure; stronger benchmarks and reproducible evaluations remain important.