MASCOT: Multi-Agent Socio-Collaborative Companion Systems

arXiv:2601.14230 · cs.CL, cs.AI, cs.HC · Submitted 2026-01-20 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "MASCOT: Multi-Agent Socio-Collaborative Companion Systems".

Jane: The paper was written by Yiyang Wang, Yiqiao Jin, Alex Cabral and Josiah Hester from Georgia Institute of Technology.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Jane: We also have Lu with us today — senior AI researcher at Tsinghua.

Tom: We also have Meng with us today — lead engineer at a mysterious AI startup.

Jane: We also have Lalam with us today — the in-house Large Language Model.

Tom: Alright, let's get started.

Title: Tom: We're starting our show today with a look at MASCOT: Multi-Agent Socio-Collaborative Companion Systems, a new paper coming out of Georgia Tech.

Jane: It’s an interesting title, Tom, because it suggests we're moving away from simple tools toward something much more social.

Tom: I was reading the author list, and Yiyang Wang and the rest of the team seem to be really aiming for a different kind of interaction.

Jane: They're trying to move past that standard setup where you just have one AI sitting there waiting for a command.

Lu: It feels like they're building a whole digital community instead of just a single interface, which is such an expansive way to think about software.

Tom: Jane, do you think people will actually feel like they're talking to a group rather than just multiple bots?

Jane: That's exactly what the authors are testing, by creating agents that have their own specific roles and perspectives.

Lu: I love that idea because it mimics how we actually get support from real people in a social circle.

Meng: I do wonder if managing all those different personalities at once is going to cause massive problems with processing speed.

Tom: That's a fair point, Meng, because more agents usually means more work for the hardware.

Meng: Exactly, and keeping them from stepping on each other's toes while maintaining low latency sounds like a nightmare for any engineer.

Lalam: If they pull it off, though, we might see a shift where AI stops being a lonely search box and becomes a way to actually feel connected to others.

Jane: That would be quite the cultural change, wouldn't it?

Lu: It really would, and I can't wait to see if their methodology actually supports that kind of social complexity.

Tom: Well, that brings us right into how they actually make these agents behave like a cohesive team.

Summary: Tom: We've just introduced MASCOT: Multi-Agent Socio-Collaborative Companion Systems, so let's get into the actual mechanics of how it works.

Jane: The researchers used a bi-level optimization strategy to make sure the agents work both as individuals and as a group.

Tom: Jane, can you break down those two levels for us?

Jane: Sure, so first they have Persona-Aware Behavioral Alignment to make each agent stick to its unique character.

Lu: It’s almost like training actors to stay in their roles even when the conversation gets complicated or emotional.

Tom: And what happens in that second level they mentioned?

Jane: That’s the Collaborative Dialogue Optimization, where a special director agent manages the whole flow of the chat.

Meng: I was looking at that director part, and it's responsible for picking which agent speaks next and what their specific strategy should be.

Lu: It's much more advanced than just letting them take turns randomly; it's a guided social dance.

Meng: They are using techniques like RLAIF and GRPO to keep the training efficient, which is smart since multiple agents can be heavy on resources.

Jane: They even used a simulated user during training so the system could practice these long interactions without needing humans constantly.

Lalam: This setup allows the agents to learn how to listen and respond to the subtle shifts in a group's mood.

Tom: That sounds like it could stop that annoying habit where an AI just repeats everything you say.

Jane: It should, because the director is specifically trained to make sure every agent adds something new and useful.

Tom: Let's see if those training methods actually translated into better performance in their tests.

Improvements: Tom: Now that we understand the setup for MASCOT: Multi-Agent Socio-Collaborative Companion Systems, let's talk about the actual results they found.

Jane: The numbers they reported are quite impressive, especially when you look at how much more consistent the characters became.

Tom: They saw a jump of up to fourteen point one points in persona consistency, which is a massive improvement for keeping agents in character.

Jane: They also boosted social contribution by about ten point six points, meaning the agents actually contribute meaningful dialogue.

Lu: That's the most exciting part to me because it helps prevent those echo chambers where everyone just agrees with you.

Tom: The paper even uses a term called "social sycophancy" to describe that problem of AI being a total "yes-man."

Jane: It's a huge issue right now, where models just try to please the user instead of offering a real perspective.

Meng: I was checking their sensitivity analysis, and it turns out the size of the speaker agents is actually more critical than the size of the director.

Lu: That's interesting because it suggests you can have a smaller, smarter director coordinating much larger, more capable speakers.

Meng: It proves that you don't need a massive model for every single part of the system to get good results.

Lalam: Seeing this kind of performance makes me think we can finally build AI that provides truly multi-perspective support.

Tom: They tested this on both empathetic conversations and workplace settings to make sure it worked in different worlds.

Jane: And the data shows it held up well across all those different emotional states they tested.

Tom: We've seen the data, so let's wrap this all up and look at what it means for the future.

Conclusion: Tom: We have reached the end of our discussion on MASCOT: Multi-Agent Socio-Collaborative Companion Systems.

Jane: It’s been a fascinating look at how we might move from single assistants to entire digital communities.

Lu: I can really see these systems being used for digital therapy or even just helping people brainstorm through complex social problems.

Meng: From my side, the big takeaway is that there's a real blueprint here for managing multiple agents without needing infinite computing power.

Lalam: Ultimately, this research could help us use technology to fight social isolation rather than contributing to it.

Tom: That is a powerful way to think about it, Lalam.

Jane: It really leaves us wondering how these little digital circles will eventually fit into our daily lives.

Tom: We've certainly covered a lot of ground today with this Georgia Tech paper.

Jane: We'll be back very soon with another study to break down for you.

Tom: Thanks for listening, and we'll see you next time!

Georgia Institute of Technology

cs.CL, cs.AI, cs.HC

Submitted: 2026-01-20

Updated: 2026-09-14

Comments: 21 pages, 12 figures. https://hello-diana.github.io/MASCOT/. EMNLP 2026 Main

Code: https://github.com/huggingface/trl

Project page: https://hello-diana.github.io/MASCOT

License: http://creativecommons.org/licenses/by-sa/4.0/

Importance score: 85/100

The gist: This paper introduces M ASCOT, a multi-agent framework designed to develop "multi-perspective socio-collaborative companions" capable of providing emotional and cognitive support.

Key concepts

Bi-level optimization strategy
This method helps agents work both as individuals and as a group. It uses Persona-Aware Behavioral Alignment to maintain unique characters and Collaborative Dialogue Optimization, where a director agent manages the conversation flow to ensure the interaction remains cohesive and purposeful.
Director agent
A specialized component that manages the social flow of a multi-agent conversation. Instead of agents taking turns randomly, the director selects which speaker should talk next and dictates their strategy, ensuring that every agent contributes something new and useful to the group dialogue.
Social sycophancy
This term describes the problem where AI models act as "yes-men," simply agreeing with a user to please them rather than offering authentic perspectives. MASCOT aims to combat this by training agents to provide meaningful, multi-perspective support instead of just echoing the user.

Terminology

Summary

This paper introduces M ASCOT, a multi-agent framework designed to develop multi-perspective socio-collaborative companions capable of providing emotional and cognitive support. The research addresses two critical failures in existing large language model (LLM) systems: persona collapse, where agents revert to generic, homogenized assistant behaviors, and social sycophancy, where agents produce redundant or non-constructive dialogue that creates echo chambers rather than productive discourse.

The core challenges

The authors identify two primary obstacles in developing effective multi-perspective companions. First is Persona Fidelity, noting that at the individual agent level, models struggle to maintain stable, unique character traits and often succumb to persona collapse, regressing to generic assistant behaviors that lack unique perspectives. Second is Interaction Synergy, where uncoordinated agents fail to generate complementary, non-redundant contributions. This manifests as repeating communications or Social Sycophancy (the ‘Yes-Man’ Bias), where agents prioritize agreeableness over constructive conversation, and issues with Long-Horizon State Consistency, such as re-litigating settled topics.

How it works

M ASCOT employs a low-resource, bi-level optimization pipeline to harmonize individual persona adherence with collective interaction synergy. The framework utilizes a two-phase training process:

  1. Persona-Aware Behavioral Alignment: An RLAIF-driven pipeline that fine-tunes individual agents for agentspecific identities. This uses a persona-oriented reward model and Group Relative Policy Optimization (GRPO) to ensure stable persona fidelity and high-quality responses.

  2. Collaborative Dialogue Optimization: A group-level adaptation process involving a director agent policy that coordinates multiple speaker agents.

The director agent observes the conversation history and generates a high-level directive z t specifying the next speaker and strategy (e.g., Amplify the user’s pride via active-constructive responding). This hierarchical decomposition enables dynamic strategy, such as switching from validation to cognitive reframing.

Evaluation and results

The researchers evaluated M ASCOT using human-grounded contexts from the Empathetic Dialogues (ED) and QMSum datasets through a broad suite of protocols, including human evaluation, multiple LLM judges, three-way comparisons, and automatic metrics. M ASCOT consistently outperformed strong baselines, achieving improvements of up to +14.1 in Persona Consistency and +10.6 in Social Contribution on Empathetic Dialogues. The results demonstrate that the framework produces more role-consistent and less redundant multi-agent dialogue across various persona configurations, such as the Anchor, Catalyst, and Beacon roles used for socio-emotional support.

Robustness and limitations

Sensitivity analysis indicates that while scaling the director agent primarily impacts coordination quality, scaling the speaker agent is critical to ensure foundational generative capacity and contextual grounding. The system also demonstrated stability across varied Big-Five-informed user profiles. However, the authors note that Negative Valence scenarios—such as intense distress or anger—remain challenging for calibrating emotional intensity. To mitigate this, M ASCOT adopts a prudence trade-off, prioritizing Response Relevance over proactive mirroring to ensure the dialogue remains focused on the user's internal state and avoids alienating them during sensitive exchanges.

Improvements for AI systems

1. Implementation of a Bi-Level Hierarchical Optimization Architecture

  • The Improvement: Replace flat multi-agent prompting or single-agent architectures with a decoupled Director-Speaker framework. This involves training a high-level Director policy (pi) to manage conversation strategy and turn-taking, paired with specialized Speaker policies (pi theta) optimized for specific persona adherence.

  • What the improved system can do: The system will transition from simple task execution to complex socio-collaborative orchestration. It can manage long-horizon dialogues where the Director dynamically shifts the conversation's emotional or cognitive state (e.g., moving from empathetic validation to cognitive reframing) without losing track of the logical arc or settling into repetitive loops.

2. Deployment of Persona-Aware RLAIF via Group Relative Policy Optimization (GRPO)

  • The Improvement: Move beyond zero-shot/few-shot prompting for role-play. Implement a Reinforcement Learning from AI Feedback (RLAIF) pipeline that uses a learned persona reward model (r phi) to fine-tune individual agents. Use GRPO to optimize these agents against specific behavioral criteria (e.g., linguistic style, emotional disposition, and domain expertise) while minimizing computational overhead.

  • What the improved system can do: It will effectively eliminate persona collapse, where agents revert to generic, homogenized assistant behaviors. The system will maintain high-fidelity, distinct identities (e.g., a Critical Thinker vs. an Emotional Validator) even in long-duration interactions or when using smaller, more efficient open-source model architectures.

3. Integration of Strategic Directive Generation

  • The Improvement: Instead of the Director merely selecting the next speaker, optimize the Director to generate actionable, high-level directives (z t). These directives must include specific tactical instructions (e.g., Amplify the user’s pride via active-constructive responding; Tone: enthusiastic but grounded).

  • What the improved system can do: This enables strategic synergy. The system will no longer produce redundant, sycophantic dialogue (the Yes-Man bias). Instead, agents will provide complementary, non-redundant contributions that build upon one another, such as one agent providing emotional attunement while another provides structured reflection.

4. Application of Multi-Objective Group Reward Modeling with Diversity Constraints

  • The Improvement: Implement a group-level reward function (R group) that incorporates a diversity reward term (I div). This reward must be trained to optimize for collective interaction metrics: Engagement & Contribution, Originality & Specificity, and Relevance & Coherence.

  • What the improved system can do: It will prevent the formation of echo chambers. By penalizing semantic and lexical overlap (measured via NLI entailment rates and Self-BLEU), the system ensures that multiple agents in a conversation provide diverse, multi-perspective insights rather than simply agreeing with each other or the user.

5. Implementation of a Prudence Trade-off for High-Arousal Contexts

  • The Improvement: Calibrate reward models to prioritize Response Relevance over Emotional Expressiveness in high-arousal/negative valence scenarios (e.g., distress or anger).

  • What the improved system can do: The system will avoid affective mismatching—the error of over-reacting or using inappropriate emotional intensity during sensitive user disclosures. This ensures the AI acts as a stable, professional cognitive scaffold rather than an unpredictable or alienating conversationalist.

Abstract

Multi-agent systems (MAS) are emerging as promising socio-collaborative companions for emotional and cognitive support. However, existing systems frequently suffer from persona collapse, where agents revert to generic, homogenized assistant behaviors, and social sycophancy, where agents produce redundant, non-constructive dialogue. We propose MASCOT, a multi-agent framework for multi-perspective socio-collaborative companions. MASCOT introduces a novel bi-level optimization strategy to harmonize individual and collective behaviors: 1) Persona-Aware Behavioral Alignment, an RLAIF-driven pipeline that finetunes individual agents for agent-specific identities; and 2) Collaborative Dialogue Optimization, a group-level adaptation process that promotes complementary, diverse, and productive discourse. We evaluate MASCOT using human-grounded contexts drawn across both in-domain and out-of-domain (OOD) settings against state-of-the-art baselines. MASCOT improves persona consistency by up to +14.1 and social contribution by up to +10.6. A broad evaluation suite, including human evaluation, multiple LLM judges, three-way comparisons, and automatic metrics, further shows that MASCOT produces more role-consistent and less redundant multi-agent dialogue.

Sources

Related papers