Adaptive Switching Between Self-Mechanisms, Mutual Mechanisms, and Goal Mechanisms in Multiple Large Language Models: A Goal-Directed Framework for Dynamic Intelligence Coordination
Year: 2026
Type: preprint
Abstract: Large language models (LLMs) can improve task performance through individual processing mechanisms, including self-reflection, self-evaluation, self-correction, self-verification, and self-planning. Multiple LLMs can also interact through mutual evaluation, collaborative planning, cross-model verification, mutual correction, and distributed reasoning. However, neither individual processing nor multi-model collaboration is necessarily sufficient to ensure that a system consistently pursues its intended objectives. A system may produce locally convincing responses while deviating from the task goal, optimize an easily measured intermediate objective at the expense of the final outcome, or continue generating and revising outputs without making meaningful progress. This paper proposes Adaptive Switching between Self-Mechanisms, Mutual Mechanisms, and Goal Mechanisms (ASSM-MM-G), a conceptual framework that integrates individual-model processing, cross-model interaction, and explicit goal-oriented control. The framework represents goals as operational structures containing desired outcomes, constraints, priorities, intermediate milestones, evaluation criteria, and termination conditions. A goal-aware controller dynamically selects among individual processing, multi-model collaboration, goal analysis, goal decomposition, goal evaluation, goal revision, and goal-progress monitoring. Its decisions depend on the current task state, estimated goal satisfaction, uncertainty, model disagreement, expected improvement, and computational cost. The framework distinguishes changes in processing mechanisms from changes in the objectives those mechanisms serve, allowing goal-directed behavior to remain stable while processing strategies adapt. It also permits goal revision when requirements change, evidence invalidates assumptions, or objectives conflict, subject to explicit constraints on permissible goal changes. This paper defines the framework, formalizes its decision problem, presents a reference algorithm, identifies failure modes, and proposes research questions, hypotheses, and an experimental protocol. The framework is a research proposal rather than an empirically validated system. Its potential contribution is the investigation of whether goal representation and goal-state feedback can improve adaptive mechanism selection beyond systems that select individual or collaborative processing mechanisms without an explicit, dynamically evaluated goal structure.
Author: Zhang, Jincheng
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7222807041
