AI Researchers Propose Mental Imagery for Dialogue Grounding

A new arXiv preprint (2604.21144) introduces a method to improve situated dialogue in conversational agents by using machine mental imagery to represent common ground. The authors identify a failure mode called 'representational blur,' where fine-grained distinctions between similar entities collapse into interchangeable textual descriptions, creating an illusion of grounding. Inspired by human mental imagery and multimodal models, they propose giving agents an analogous ability to construct depictive internal representations to preserve shared context beyond immediate context windows. The research targets a key limitation of current agents that struggle to maintain reliable shared context over time.