Adaptive Switching Between Self-Planning and Collaborative Planning for Multiple Large Language Models

Abstract

Large language models (LLMs) are increasingly employed as autonomous agents capable of decomposing tasks, generating action sequences, evaluating alternatives, and coordinating with other agents. However, existing planning strategies face a fundamental trade-off. Self-planning allows an individual model to generate and revise plans with relatively low communication overhead, but it may preserve individual reasoning errors, overlook alternative strategies, and become trapped in a limited planning trajectory. Collaborative planning enables multiple models to exchange information, compare alternatives, and construct more comprehensive plans, but it can introduce communication costs, coordination delays, redundant reasoning, and correlated errors. This paper proposes Adaptive Self–Collaborative Planning (ASCP), a conceptual framework that dynamically switches among self-planning, collaborative planning, and optional transition modes according to task characteristics, planning uncertainty, inter-model disagreement, estimated plan quality, execution feedback, and computational cost. Rather than requiring every task to follow a fixed planning procedure, ASCP treats the selection of planning mode as a decision-making problem in its own right. A meta-controller evaluates the current planning state and selects an appropriate mode, while a shared task representation maintains consistency across transitions. The framework incorporates plan evaluation, disagreement-sensitive collaboration, execution-based replanning, and cost-aware mode selection. A formal objective, an adaptive switching algorithm, research hypotheses, and a reproducible experimental design are presented. The framework is intended to establish a testable research direction for balancing planning autonomy and multi-model coordination. Its effectiveness remains an empirical question requiring systematic comparison with fixed-mode baselines.