When Language Hides Causes: On Causal Representation Problems in LLMs

This paper draws a conceptual distinction between the representation of an input and the representation of its cause. It focuses on the latter to systematically examine the epistemic challenges faced by any agent that develops representations of the causes of its inputs—challenges that, by extension, concern any model that implicitly constructs a world model from its input data. We argue that these problems manifest saliently in the case of Large Language Models (LLMs), but that they do not constitute an in-principle limitation of such systems. On the contrary, our main thesis is that current obstacles to reasoning and generalization in LLMs arise, at least in part, not from the absence of human-like multimodal embodiment, but from the structure of human linguistic practices that govern the data on which these models are trained.