MIRROR: Cognitive Inner Monologue Between Conversational Turns for Persistent Reflection and Reasoning in Conversational LLMs
MIRROR: Cognitive Inner Monologue Between Conversational Turns for Persistent Reflection and Reasoning in Conversational LLMs
Human intelligence relies on inner monologue to process complex information through simultaneous reflection, memory retrieval, and response formulation. We introduce MIRROR (Modular Internal Reasoning, Reflection, Orchestration, and Response), a cognitive architecture that systematically implements these parallel reasoning capabilities in large language models. MIRROR operates as a unified system with two distinct functional layers: the Thinker and the Talker. The Thinker encompasses: (1) the Inner Monologue Manager, coordinating reasoning threads across cognitive dimensions (Goals, Reasoning, and Memory); and (2) the Cognitive Controller, synthesizing these threads into a coherent internal narrative maintained across conversation turns. The Talker component then leverages this integrated narrative for context-aware responses. Evaluated on the CuRaTe benchmark--testing personalized dialogue with safety-critical constraints, conflicting preferences, and multi-turn consistency--LLMs utilizing the MIRROR architecture achieve up to 156% relative improvement in critical safety scenarios involving three persons with conflicting preferences, maintaining an average accuracy of ~>80% on all scenarios. Across scenario-specific comparisons, GPT-4o, Gemini 1.5 Pro, Claude 3.7 Sonnet, Llama 4 variants, and Mistral 3 variants with the MIRROR architecture outperformed baseline models by 21% on average (15.5 percentage points absolute). MIRROR directly addresses three critical LLM failure modes: sycophancy, attentional deficits to critical information, and inconsistent prioritization of conflicting constraints. This work bridges cognitive science and AI by implementing modular internal reasoning inspired by human cognition, creating a persistent internal model that significantly enhances multi-turn conversation capabilities.
Nicole Hsing
计算技术、计算机技术
Nicole Hsing.MIRROR: Cognitive Inner Monologue Between Conversational Turns for Persistent Reflection and Reasoning in Conversational LLMs[EB/OL].(2025-05-31)[2025-07-17].https://arxiv.org/abs/2506.00430.点此复制
评论