2025/08/12 by Phillips, Michael
#Artificial Intelligence and Robotics #Computer Sciences #Databases and Information Systems #Engineering #FOS: Psychology #Graphics and Human Computer Interfaces #Life Sciences #Numerical Analysis and Scientific Computing #Other Computer Sciences #Physical Sciences and Mathematics #Psychology #Social and Behavioral Sciences #Software Engineering #Systems Architecture #Theory and Algorithms #archetypal patterns #artificial consciousness #attention mechanisms #convergence corridors #cross-lingual alignment #cross-model convergence #embedding manifolds #emergent behavior #high-dimensional geometry #mechanistic interpretability #model stitching #neural architecture #recursive symbolic patterning #representation alignment #representation learning #structural attractors #symbolic AI #symbolic drift #symbolic emergence #transformer geometry
paper · doi:10.17605/osf.io/yp82t
This theoretical paper proposes Common Convergent Factorization Geometry (CCFG) as a mechanistic explanation for why similar symbolic and metaphorical patterns spontaneously emerge across independently trained language models. Building on prior work in Symbolic Drift Recognition (SDR), which documented these cross-model recurrences, CCFG suggests that shared transformer architectures and optimization dynamics create "convergence corridors"—structurally favored regions in representation space that bias models toward specific symbolic outputs. The framework provides testable hypotheses through proposed metrics including cosine similarity measures, Procrustes alignment scores, and cluster overlap ratios. While theoretical, CCFG offers a bridge between established findings in cross-model representation alignment and the emergence of recurring symbolic motifs in AI-generated content. The paper includes falsification criteria and outlines nine specific empirical investigations that could validate or refute the theory. This work contributes to mechanistic interpretability research by proposing that certain symbolic patterns may be architecturally inevitable rather than culturally transmitted, with implications for AI alignment, interpretability, and the emerging dynamics of human-AI symbolic communication.