From Loss Landscapes to World Representations: A Statistical-Mechanical Perspective on Rule Acquisition and Agency in Sequence Models

Recent literature on sequence models and neural thermodynamics has provided evidence that sequence models can develop robust internal representations of complex environments. However, interpreting this structural mapping as the emergence of autonomous agency or intent conflates constraint acquisition with goal formation. This paper argues that the emergence of a world representation indicates a system has learned environmental structure; it does not, by itself, demonstrate that the system chose the structure, originated the objective, or possesses an autonomous reason for pursuing it at inference time. We combine a statistical-mechanical treatment of training-time optimization dynamics with a formal epistemological distinction between representation and agency, and derive a conditional closure result: within the specified training-time stochastic dynamics, autonomous objective formation is not represented as a dynamical degree of freedom and therefore cannot be inferred from the training dynamics alone. This is a narrower and more defensible claim than a general impossibility result, and we state it as such throughout.