Self-Evolution Is Not Enough: Toward a Historically Materialist AI for Scientific Problem Solving

Recent work on self-improving agents, evolutionary search, autonomous scientific discovery, and open-ended artificial intelligence has substantially expanded the range of processes through which an artificial system can revise hypotheses, programs, strategies, tools, and even parts of its own architecture. Yet self-improvement does not by itself amount to historical self-evolution. Scientific problems do not exist in an abstract and temporally uniform solution space. What can be formulated, investigated, tested, and established at a given moment depends upon historically accumulated concepts, evidence, instruments, methods, forms of labor, institutions, infrastructures, and relations of knowledge production. This paper develops a philosophical account of artificial intelligence for scientific problem solving from this observation. Drawing on Hegelian dialectics, historical materialism, and an analysis of the unequal structural roles of contradictions within developing systems, it treats problem solving as the transformation of a historically constituted problem situation rather than as the optimization of candidate answers under a fixed objective. The central proposal is that a genuinely self-evolutionary scientific AI would need to reconstruct the material and relational conditions under which a problem has become possible, identify contradictions that constrain or organize further development, distinguish structurally central contradictions from derivative ones, compare interventions by the research states they make reachable, and revise its own representation of the problem after those interventions alter the conditions of inquiry. On this view, an apparently minor intermediate problem may be scientifically important because solving it creates data, methods, instruments, distinctions, institutions, or coordination capacities through which a previously inaccessible problem becomes researchable. The paper therefore distinguishes conceptual possibility, present researchability, and epistemic establishability, and introduces the problem of epistemic stage-skipping in AI-assisted research. It further argues that methods and evaluative regimes are themselves historically situated components of scientific production. A method or criterion that once enables inquiry can become restrictive as the material, conceptual, and institutional capacities of research develop. Under such conditions, stronger optimization against an inherited evaluator can deepen epistemic misalignment rather than resolve it. Historically self-evolutionary AI must therefore be able to diagnose when methods or evaluators have exceeded their domain of adequacy and to transform them through a documented lineage that preserves what remains valid. The resulting account is neither an engineering architecture nor a claim that Hegelian or Marxian concepts can be directly converted into algorithms. It is a philosophical design orientation for long-horizon artificial intelligence whose task is not merely to produce better answers, but to participate in the historical transformation of what can responsibly be asked, known, tested, and solved.