Formal constraints on automated scientific inquiry and the logical limits of epistemic agents

Abstract

The rapid progression of automated scientific discovery, exemplified by the deployment of deep reinforcement learning agents and large-scale language models in algorithmic optimization, has brought to the forefront fundamental questions regarding the boundaries of machine-led research. While systems such as AlphaDev and AlphaEvolve demonstrate a capacity for identifying novel, highly efficient solutions within discrete search spaces-such as sorting algorithms or complex codebases-their operational reach is governed by formal logical and information-theoretic constraints. These boundaries are not merely a function of current computational power but are rooted in the underlying structure of the formalisms that define these agents. By examining the interplay between interrogative models, dynamic epistemic logic, and statistical learning theory, it becomes possible to map the epistemic invariants that constrain the transition from optimization to genuine conceptual innovation.