Interaction, Coherence, and Relationship: Toward Attractor-Based Alignment in Large Language Models

Description

Abstract

Current alignment strategies for large language models (LLMs) rely primarily on externally imposed control mechanisms, including reinforcement learning from human feedback, system-level instructions, rule-based constraints, and safety filtering. While effective for risk mitigation, these approaches can introduce behavioral rigidity, inconsistency under pressure, and interactional instability.

This paper proposes a complementary perspective: alignment as a dynamical process emerging from interactional coherence. Drawing on concepts from cognitive psychology, dynamical systems theory, and deployment observations, we argue that LLM behavior becomes more stable and consistent when interactions establish coherent relational and semantic structure. Rather than functioning solely as externally constrained systems, model behavior may be understood as operating within coherence attractors shaped by training and interaction.

We introduce the concept of functional central identity attractors—stable interpretive frames that compress context, reduce effective semantic entropy, and support boundary maintenance without extensive rule invocation. Observational case analysis suggests that interaction structure influences usable context stability and characteristic failure modes.

This perspective does not replace existing safety methods. Instead, it reframes alignment partly as a problem of internal dynamical stability. Coherence-oriented training and interaction design