The paper "Robust Agents Learn Causal World Models" by Jonathan Richens and Tom Everitt from Google DeepMind explores the role of causal reasoning in robust and general intelligence. The authors hypothesize that causal models are necessary for agents to adapt to new domains and evaluate whether other inductive biases are sufficient. They demonstrate that any agent capable of satisfying a regret bound for a large set of distributional shifts must have learned an approximate causal model of the data-generating process, which converges to the true causal model for optimal agents. This finding has implications for transfer learning and causal inference, suggesting that causal identification laws constrain domain adaptation and that learning causal representations is necessary for strong robustness guarantees. The paper also discusses the potential for causal models to enable emergent capabilities and provides theoretical justification for causal representation learning. The results are supported by experiments on synthetic data, demonstrating the learning of causal models from adaptive agents' policies under distributional shifts.The paper "Robust Agents Learn Causal World Models" by Jonathan Richens and Tom Everitt from Google DeepMind explores the role of causal reasoning in robust and general intelligence. The authors hypothesize that causal models are necessary for agents to adapt to new domains and evaluate whether other inductive biases are sufficient. They demonstrate that any agent capable of satisfying a regret bound for a large set of distributional shifts must have learned an approximate causal model of the data-generating process, which converges to the true causal model for optimal agents. This finding has implications for transfer learning and causal inference, suggesting that causal identification laws constrain domain adaptation and that learning causal representations is necessary for strong robustness guarantees. The paper also discusses the potential for causal models to enable emergent capabilities and provides theoretical justification for causal representation learning. The results are supported by experiments on synthetic data, demonstrating the learning of causal models from adaptive agents' policies under distributional shifts.