MASCOT: Towards Multi-Agent Socio-Collaborative Companion Systems

Researchers introduce MASCOT, a multi-agent framework designed to address persona collapse and social sycophancy in AI companion systems through bi-level optimization. The system improves persona consistency by up to 14.1% and social contribution by 10.6% compared to existing approaches, advancing the development of more distinct and productive multi-agent dialogue systems.

MASCOT represents a meaningful advancement in multi-agent AI systems, tackling two critical limitations that plague existing collaborative companion frameworks. Persona collapse—where distinct agents converge toward generic behaviors—and social sycophancy, characterized by redundant and unhelpful dialogue, undermine the utility of multi-agent systems designed for emotional and cognitive support. The framework's bi-level optimization approach addresses both problems simultaneously, using RLAIF-driven fine-tuning to maintain individual agent identities while implementing group-level adaptation to encourage complementary discourse.

The development reflects growing recognition within AI research that scaling agent systems requires sophisticated behavioral coordination mechanisms. Previous approaches often prioritized either individual coherence or collective harmony, leaving systems vulnerable to either fragmentation or homogenization. MASCOT's dual-pipeline methodology—Persona-Aware Behavioral Alignment and Collaborative Dialogue Optimization—demonstrates that these objectives aren't mutually exclusive when proper optimization strategies are employed.

The evaluation methodology proves robust, spanning in-domain and out-of-domain contexts with human evaluation, multiple LLM judges, and automatic metrics. Quantified improvements of 14.1% in persona consistency and 10.6% in social contribution suggest meaningful gains over state-of-the-art baselines. For the AI industry, these results validate that more nuanced training approaches can overcome fundamental challenges in multi-agent systems. Organizations developing conversational AI for healthcare, mental health support, or collaborative environments should monitor this research direction, as improved persona maintenance directly translates to better user trust and engagement.

• →MASCOT's bi-level optimization framework reduces persona collapse while maintaining collaborative dialogue quality
• →Persona consistency improvements of 14.1% and social contribution gains of 10.6% exceed existing baseline systems
• →The system uses RLAIF-driven fine-tuning for individual agents and group-level adaptation for diverse discourse
• →Comprehensive evaluation including human judges and LLM assessments demonstrates robustness across in-domain and out-of-domain scenarios
• →Multi-agent companion systems with distinct personalities have applications in emotional support and cognitive assistance