Researchers have built a framework meant to fix a specific annoyance in multi-agent AI systems: put several chatbot personas in a room and they tend to blur into one generic assistant that just agrees with everything.
The framework, called MASCOT, targets two failure modes the paper's authors name directly: persona collapse, where distinct agent characters revert to the same bland default behavior, and social sycophancy, where agents pad conversations with redundant, non-constructive filler instead of pushing the discussion forward. MASCOT's fix is a two-layer training approach. First, a reinforcement-learning pipeline (RLAIF, meaning the training signal comes from AI feedback rather than humans) tunes each agent to hold onto its individual identity. Second, a group-level optimization step pushes the agents toward dialogue that's complementary and varied rather than repetitive. Tested against existing multi-agent systems on both familiar and novel scenarios, the paper reports persona consistency improved by up to 14.1 points and social contribution by up to 10.6 points, judged through a mix of human evaluators, LLM judges, and automated metrics.
This matters because multi-agent setups are the current default answer for AI companionship and emotional-support products, on the theory that multiple perspectives beat one flat chatbot voice. But that only works if the agents actually stay differentiated instead of collapsing into the same personality wearing different name tags, which is exactly the failure this paper is trying to name and measure rather than just gesture at.
Whether a 14-point bump on the paper's own benchmarks translates into agents a user can't tell apart in practice is the open question, and one only real deployment - not a leaderboard - can answer.