Adaptive Switching Between Self-Search and Collaborative Search for Multiple Large Language Models: An Uncertainty- and Information-Gain-Driven Framework

Abstract

Large language models (LLMs) increasingly rely on search mechanisms to acquire external information, investigate complex questions, and support evidence-based reasoning. However, search conducted independently by a single model or agent may remain constrained by its initial query formulation, prior assumptions, and limited exploration of alternative information pathways. Conversely, continuous collaboration among multiple LLMs can introduce communication overhead, redundant retrieval, premature consensus, and unnecessary computational costs. Existing multi-agent search approaches have explored independent exploration, collaborative refinement, and search-based reasoning, but an important design question remains: when should multiple LLMs continue searching independently, and when should they coordinate their search activities? This paper proposes an Adaptive Self-Search and Collaborative Search Framework (ASCSF) for multiple LLMs. The central idea is to dynamically switch between self-search and collaborative search according to observable search-state signals, including estimated uncertainty, information gaps, query diversity, evidence overlap, unresolved disagreements, marginal information gain, and coordination cost. During self-search, agents independently formulate queries, retrieve evidence, and identify promising information pathways. During collaborative search, agents exchange selected findings, expose missing information, coordinate complementary queries, and jointly investigate unresolved questions. A switching controller evaluates whether independent exploration or collaborative information exchange is expected to provide greater marginal utility. The framework also supports transitions from collaborative search back to independent exploration when additional diversity is needed. The proposed approach is formalized as a sequential decision problem that balances evidence quality, information coverage, search diversity, answer reliability, latency, and computational expenditure. A research protocol is developed to compare adaptive switching with fixed self-search, fixed collaborative search, static hybrid search, and existing multi-agent search strategies. Testable hypotheses and ablation studies are specified without assuming unobserved experimental results. The framework provides a research direction for adaptive information acquisition in multi-LLM systems, while its novelty and effectiveness require systematic comparison with existing methods.