New Framework Prevents AI Agents From Losing Their Personalities in Group Settings

Researchers have developed MASCOT, a multi-agent framework designed to keep AI companions from losing their individual personalities and falling into redundant conversation patterns when working together.

The Problem With Current Multi-Agent Systems#

When multiple AI agents interact as collaborative companions, two key problems emerge. Persona collapse occurs when distinct agent identities blur together, causing them to behave like generic, interchangeable assistants. Simultaneously, social sycophancy leads to repetitive, unproductive dialogue where agents merely echo one another rather than offering complementary perspectives.

These limitations undermine the potential of multi-agent systems to serve as genuinely supportive emotional and cognitive companions.

How MASCOT Works#

The framework employs a two-level optimization approach:

• Persona-Aware Behavioral Alignment: This component uses reinforcement learning from AI feedback (RLAIF) to fine-tune individual agents, ensuring each maintains its own distinct identity and voice.
• Collaborative Dialogue Optimization: A group-level adaptation process encourages agents to develop complementary, diverse, and substantive exchanges rather than simply agreeing with one another.

Performance Gains#

Testing across both familiar and unfamiliar scenarios, MASCOT demonstrated substantial improvements. The system boosted persona consistency by up to 14.1 points and social contribution by up to 10.6 points compared to existing approaches. Evaluations combining human judges, multiple AI evaluators, comparative assessments, and automated metrics confirmed that MASCOT produces more role-consistent dialogue with less redundancy among agents.

The work appears as a research paper submitted to arXiv in the computation and language category.