MIRROR: Cognitive Inner Monologue Between Conversational Turns for Persistent Reflection and Reasoning in Conversational LLMs
MIRROR implements human-like inner monologue in AI through a cognitive architecture with parallel reasoning capabilities. Achieves up to 156% improvement in safety-critical scenarios and 21% average performance boost across GPT-4o, Claude 3.7, Llama 4, and more.
156%+
Safety improvement
21%
Avg. performance boost
80%+
Accuracy maintained
Key Contributions
- Novel cognitive architecture inspired by human inner monologue
- Systematic implementation of parallel reasoning capabilities
- Addresses critical LLM failure modes: sycophancy, attentional deficits, inconsistent prioritization
- Bridges cognitive science and AI through modular internal reasoning