MASCOT

Multi-agent systems are increasingly used as companions for emotional and cognitive support, but they often drift into generic assistant behavior (persona collapse) or pile on redundant, agreeable dialogue (social sycophancy). MASCOT is a framework for multi-perspective socio-collaborative companions that addresses both problems with a bi-level optimization strategy: a persona-aware behavioral alignment pipeline that gives each agent a distinct identity, and a collaborative dialogue optimization step that pushes the group toward complementary, productive discourse.

My role. I led this project — formulating the framework, building the system, and designing the evaluation. Across in-domain and out-of-domain settings, MASCOT improves persona consistency by up to +14.1 and social contribution by up to +10.6, verified through human evaluation, multiple LLM judges, and automatic metrics.