Adaptive Switching Between Self-Evidence and Mutual-Evidence Mechanisms in Multiple Large Language Models: A Unified Framework for Dynamic Evidence Coordination

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Large language models (LLMs) generate answers using learned representations and contextual information, but their outputs may contain unsupported claims, inaccurate statements, incomplete reasoning, and inconsistencies with available sources. Evidence-based processing can help identify and reduce these problems through evidence retrieval, claim verification, source evaluation, and contradiction analysis. These operations can be conducted within an individual model's processing workflow or through interactions among multiple models. However, neither individual evidence assessment nor unconditional multi-model evidence coordination is necessarily optimal for every task. Individual assessment may preserve computational efficiency but overlook unsupported assumptions, whereas cross-model evidence assessment may introduce redundant judgments, correlated errors, communication overhead, and false consensus. This paper proposes Adaptive Switching between Self-Evidence and Mutual-Evidence Mechanisms (ASSM-ME), a conceptual framework for dynamically selecting and combining individual-model and cross-model evidence-processing mechanisms. Self-evidence mechanisms enable an individual model to identify claims, retrieve or inspect available evidence, evaluate source relevance, detect unsupported statements, and revise its output. Mutual-evidence mechanisms enable multiple models to independently assess evidence, exchange source assessments, compare claim-level judgments, identify contradictions, and coordinate evidence-based revisions. A state-aware controller considers evidence sufficiency, source reliability, claim importance, model disagreement, evidence independence, residual uncertainty, historical mechanism effectiveness, and computational cost when selecting the next mechanism. The framework supports transitions among individual evidence assessment, collaborative evidence analysis, cross-model verification, evidence-guided correction, and evidence-based termination. A cost-sensitive decision objective, reference algorithm, implementation architecture, research hypotheses, and experimental evaluation protocol are presented. The framework is a proposed research design rather than an empirically validated system. Its potential contribution is the formulation of evidence-mechanism selection as a sequential control problem spanning heterogeneous individual and collaborative evidence operations. Its novelty and practical value require comparison with existing fact-verification, retrieval-augmented generation, and adaptive multi-agent methods, followed by controlled empirical evaluation.