EVINCE: Optimizing Adversarial LLM Dialogues via Conditional Statistics and Information Theory

This paper introduces EVINCE (Entropy and Variation IN Conditional Exchanges), a framework that optimizes multi-LLM dialogues using conditional statistics and information theory. EVINCE introduces dual entropy optimization to balance perspective diversity with prior knowledge, providing quantitative measures for modulating LLM interactions. Through information-theoretic metrics and mutual information optimization, the framework demonstrates consistent improvement over single-LLM performance in applications ranging from disease diagnosis to news debiasing. We present theoretical foundations and empirical validation for this structured approach to LLM collaboration.