Computational Autopoiesis in Large Language Models: Extending 4E Cognitive Frameworks to Transformer Inference Dynamics
I formalized the ideas talked about in an earlier post and they stood up! It’s, let’s say, a pre-preprint. Welcome all feedback as usual
Here’s the abstract
We propose an extended notion of autopoiesis—computational autopoiesis—to characterize the organizational dynamics of decoder-only language models during inference. Classical autopoiesis (Maturana & Varela) requires metabolic self-production of physical components; we explicitly extend this to purely computational systems that exhibit organizational closure through self-maintenance of transient computational states. During inference, transformers continuously regenerate the computational media (activations, attention patterns, normalization states) that enable subsequent operations, while their outputs immediately re-enter as inputs in a closed operational loop. We formalize this extension, mapping classical autopoietic criteria to transformer operations with clear delineations between original concepts and our computational adaptations (Table 1). The framework yields testable signatures: (i) attention consolidation and representation divergence at speaker boundaries; (ii) power-law perturbation recovery in organizationally healthy models; (iii) uncertainty-correlated attention focusing; and (iv) routing fragmentation near externally imposed constraints. These patterns help explain alignment pathologies—inconsistent refusals, brittle templates, jailbreak sensitivity—as disruptions to organizational closure, though we acknowledge alternative mechanistic explanations. We propose complementary alignment strategies based on organizational health: epistemic scaffolding, role-aware shaping, and uncertainty-responsive generation. The framework offers the LLM community a structured interface with 4E cognitive science while maintaining clear boundaries between established theory and computational extensions.