Title: MASCOT: Towards Multi-Agent Socio-Collaborative Companion Systems

Abstract: Multi-agent systems (MAS) are increasingly recognized as valuable socio-collaborative partners for providing emotional and cognitive assistance. Nevertheless, current implementations often encounter significant challenges, notably persona collapse—where agents default to uniform, generic assistant roles—and social sycophancy, characterized by repetitive and unproductive interactions. To address these issues, we introduce MASCOT, a novel multi-agent framework designed for multi-perspective socio-collaborative companionship. MASCOT employs a distinctive bi-level optimization strategy to balance individual and group dynamics. This approach comprises two key components: first, Persona-Aware Behavioral Alignment, which utilizes an RLAIF-driven pipeline to fine-tune individual agents, ensuring distinct agent-specific identities; and second, Collaborative Dialogue Optimization, a group-level adaptation mechanism that fosters diverse, complementary, and productive conversations. We assessed MASCOT within human-grounded contexts spanning both in-domain and out-of-domain (OOD) environments, comparing it against leading state-of-the-art baselines. Our results demonstrate that MASCOT enhances persona consistency by as much as +14.1 and boosts social contribution by up to +10.6. Comprehensive evaluations, encompassing human assessments, various LLM judges, three-way comparisons, and automatic metrics, confirm that MASCOT generates dialogue that is more consistent with assigned roles and significantly less redundant than existing systems.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC