MIRROR: Cognitive Inner Monologue Between Conversational Turns for Persistent Reflection and Reasoning in Conversational LLMs

The MIRROR architecture enhances large language models by mimicking human inner monologue through modular reasoning and reflection, comprising a Thinker and Talker system that maintains an internal narrative for context-aware responses. Evaluated on safety-critical, multi-turn dialogues, models using MIRROR achieved up to 156% improvement in handling conflicting preferences and outperformed baseline models by 21% on average, addressing key failure modes like sycophancy and inconsistent constraint prioritization.