马克思主义实践认识论视角下的人机对齐路径反思OACHSSCD

Large model alignment is not merely a matter of engineering parameter tuning,but an epistemological issue concerning human-AI relations.Current alignment research mainly relies on statistical learning,symbolic manipulation,human feedback,and rule-based constraints.Although these approaches can improve the usefulness,safety,and compliance of model outputs to some extent,they tend to compress human values into preference signals that can be labeled,ranked,and optimized,thereby obscuring the practical,social,and historical dimensions of cognition.This paper introduces Marxist practical epistemology as a critical framework for reassessing the nature and limits of human-machine alignment. Marxist practical epistemology holds that knowledge arises from the active,dialectical interaction between subject and object,and is validated and developed through social practice.From this perspective,the paper draws a structural analogy between the dual dimensions of large model alignment and the dual stages of human cognition.External alignment—concerning the model's conformity to factual reality,user intent,and social norms—corresponds to the perceptual grasp of the world.Internal alignment—concerning the logical consistency and interpretability of the model's internal reasoning—corresponds to rational processing.Both dimensions must advance in parallel,just as perceptual materials and rational abstraction interact dialectically in human cognition.Alignment is thus not a static achievement but a cyclical process of practice-feedback-revision,continuously unfolding in response to changing technological conditions,social norms,and application scenarios.This cyclical character is not an incidental feature but a necessary consequence of the historical and open-ended nature of practice itself—as social conditions evolve,so too must the criteria and methods of alignment. Building on this foundation,the paper develops three practical pathways for alignment.The first concerns dynamic knowledge construction.Just as rational knowledge depends on rich and critically examined perceptual materials,large models must move beyond static training data to incorporate streaming learning,online updates,and embodied simulation.The second pathway addresses interpretable reasoning.By integrating symbolic reasoning which provides traceability and logical structure,with neural reasoning which enables large-scale pattern recognition,and by enforcing structured reasoning protocols,alignment systems can achieve both explanatory power and adaptive performance.The third pathway expands"human-in-the-loop"into bidirectional human-machine alignment:technically,through RLHF and retrieval-augmented generation;socially,by cultivating users'critical capacities to assess and selectively adopt AI-generated outputs.These pathways are not merely technical recommendations but epistemological commitments:they recognize that alignment,like human cognition itself,must remain responsive to the concrete conditions in which it operates. Despite these pathways,the paper identifies two fundamental limits.The first is linguistic representation:large models can generate coherent texts but cannot participate in the socially situated practices through which meaning is constituted.This is not a temporary shortfall but a structural asymmetry:models operate on statistical regularities,while human meaning is forged in labor,interaction,and historical life,and those domains cannot be fully captured by any corpus.The second is practical rationality:models can simulate certain forms of reasoning but lack the reflective judgment,responsibility,and value commitment that emerge from actual human activity.This limit,too,is constitutive rather than contingent.It follows from the fact that models are trained on past data,whereas human practical reason is oriented toward an open future that demands situated judgment.In response,the paper proposes two corresponding strategies:shifting from"linguistic data alignment"to"practical-contextual alignment",and shifting from"rule control"to"practice-based feedback and accountability".These strategies require scenario-specific training,stakeholder participation,value-sensitive design,and institutional mechanisms for continuous auditing and human oversight.They are consistent with China's regulatory framework for generative AI. The paper concludes that the future of alignment lies not in eliminating the differences between human cognition and machine computation,but in reconstructing practical boundaries that respect these differences.Specifically,at the linguistic level,alignment must overcome the limitations of abstract corpus-based alignment through practical-contextual alignment;at the governance level,it must overcome the inadequacy of mere rule control through practice-based feedback and accountability.Only in this way can the development of large models avoid degenerating into formalistic language fitting or mere preference accommodation,and instead achieve sustainable alignment under the normative guidance of human subjectivity,social values,and public responsibility.