Social interaction significantly impacts well-being, mental health, and cognition. Yet an estimated 1 in 6 people worldwide lack the social interactions they need, creating crises of social isolation and loneliness. As large language models (LLMs) integrate deeper into human lives, their role is shifting from passive tools to active socio-collaborative companions in affective and collaborative settings.
We propose MASCOT, a generalizable multi-agent framework for developing multi-perspective socio-collaborative companions. Unlike previous multi-agent systems optimized solely for task efficiency, MASCOT targets user-agent interaction quality, explicitly balancing individual agent persona consistency with global discourse dynamics. We introduce an efficient bi-level optimization strategy: (1) a Reinforcement-Learning-from-AI-Feedback (RLAIF) pipeline that fine-tunes individual agents for strict Persona Fidelity, and (2) a meta-agent policy guided by group-level rewards to ensure Interaction Synergy.
Extensive experiments demonstrate that MASCOT achieves significant improvements: +14.1 in Persona Consistency and +10.6 in Social Contribution compared to baseline approaches, with human evaluators preferring MASCOT in 69% of head-to-head comparisons.
Moving beyond dyadic interactions to multi-agent socio-collaborative companion systems. While single-agent support systems (left) can result in limited insights and echo chambers, multi-agent systems (right) provide diverse, balanced perspectives and foster a shared sense of community through agents with distinct roles.
At the individual agent level, models struggle to maintain stable, unique character traits. Agents often succumb to π«₯ persona collapse, regressing to π generic, homogenized assistant behaviors that lack the unique perspectives required for their specific roles.
At the collective level, uncoordinated agents fail to generate complementary, non-redundant contributions. This manifests as Social Sycophancy (the π "Yes-Man" Bias), where agents prioritize agreeableness, creating π echo chambers rather than constructive conversations.
Agents often struggle to track the logical progression of a conversation, treating each turn as a local optimization problem and losing track of the conversation's π logical arc, such as π re-litigating settled topics.
MASCOT is a multi-agent framework that optimizes the generation process through a two-phase training pipeline: (1) Persona-Aware Behavioral Alignment and (2) Collaborative Dialogue Optimization.
MASCOT optimizes multi-agent socio-collaborative companions through bi-level optimization: individual persona alignment and group dynamics coordination.
We move beyond standard few-shot prompting by explicitly aligning a base policy with persona constraints via Reinforcement Learning from AI Feedback (RLAIF).
Each candidate response is evaluated by an LLM judge (GPT-4o) against fine-grained rubrics including persona adherence and contextual appropriateness. We construct preference pairs with margin filtering and train a persona-aware reward model.
We optimize a composite reward combining persona adherence and format rewards (structured reasoning in <think> tags, concise answers), using Group Relative Policy Optimization (GRPO) for computational efficiency.
While Stage 1 optimizes individual agents for persona adherence, it does not account for group dynamics. We introduce group-level optimization to ensure cohesive community conversations.
At each turn, a director agent generates directives specifying the next speaker and a strategy (e.g., "Amplify the user's pride via active-constructive responding"); the selected agent then responds conditioned on its persona and the directive.
A unified Group Reward Model scores both the director's decision and the full conversation trajectory based on holistic criteria including global state consistency, diversity, and non-redundancy. A simulated user enables closed-loop training without a human annotation bottleneck.
A generalizable multi-agent framework that enables socio-collaborative companions in multi-perspective group conversations.
A low-resource, multi-level optimization pipeline that harmonizes individual persona adherence with collective interaction synergy.
Comprehensive human, multi-judge, and automatic evaluation across empathetic conversations and workplace meeting settings.
MASCOT consistently outperforms strong baselines across datasets and evaluation metrics, demonstrating significant improvements in both persona fidelity and interaction synergy.
Agent-specific metrics on Empathetic Dialogues across sentiment subsets (Positive / Negative / Neutral).
Group-level collaborative interaction metrics on Empathetic Dialogues.
Beyond emotional support, MASCOT generalizes to professional workplace settings using QMSum, a query-based meeting-summarization benchmark spanning Academic, Committee, and Product meetings. MASCOT coordinates four task-oriented personas to produce coherent, non-redundant collaborative discussions.
Collaborative interaction quality on QMSum across Academic, Committee, and Product meeting subsets.
We conducted an IRB-approved study with 40 annotators on Prolific. Annotators strongly preferred MASCOT in head-to-head comparisons, with the largest margins in neutral / ambiguous situations.
On the held-out ESConv dataset, MASCOT improves agent-specific quality by +10.8 and collective interaction by +5.1 over MultiAgentESC, with consistent dominance across three heterogeneous LLM judges β confirming robustness beyond single-evaluator bias.
Empathetic Dialogues: a benchmark for empathetic conversation grounded in emotional situations (~25K conversations). We categorize dialogues by sentiment valence (positive, negative, neutral) to evaluate emotional support across affective contexts.
QMSum: a query-based multi-domain meeting-summarization benchmark (1,808 queryβsummary pairs over 232 meetings) covering academic, committee, and product meetings.
Backbone: Qwen3-8B for the director, speakers, and simulated user; a Qwen3-0.6B reward model; LoRA fine-tuning (r=16, Ξ±=32) updating only 0.187% of parameters. Judges: GPT-4o, Gemma-3-27B, Phi-4.
Provides empathetic support and emotional grounding.
Offers Socratic questioning and cognitive reframing.
Provides growth-oriented positive reinforcement.
Distills transcripts into structured minutes.
Captures evidence-based decisions.
Identifies actionable items.
Identifies gaps constructively.
Zero-shot, Few-shot, Supervised Fine-tuning (SFT)
Chain-of-Thought (CoT), Self-Consistency, MultiAgentESC
We analyze the contribution of each component by systematically removing it. The persona alignment module proves most critical for maintaining individual agent quality.
Ablation across MASCOT variants.
Removing persona-aware alignment (MASCOT-P) causes the largest drop, especially in Empathetic Support Quality (72.8 β 65.1) and Consistency (76.6 β 72.3). Without explicit persona grounding, agents default to generic AI-assistant behavior.
The MASCOT-C variant with randomized speaking orders primarily affected Social Contribution and Relevance & Coherence. The collaborative module ensures each agent builds on the previous turn rather than reacting in isolation.
We vary the backbone size of the director and speaker agents to locate performance bottlenecks.
Performance sensitivity to director and speaker model size.
The director remains effective even when scaled down to 1.7B, but the speaker size is the critical bottleneck β a 0.6B speaker causes a significant performance drop.
To probe robustness across diverse users, the released pipeline can simulate users with Big Five (OCEAN) personality profiles (--enable-big-five, optionally --user-big-five-profile): the simulated user's replies and emotional trajectory are conditioned on the selected trait profile, and evaluation can be broken down per profile. See src/user/simulated_users.py in the repository.
@inproceedings{wang2026mascot,
title={MASCOT: Towards Multi-Agent Socio-Collaborative Companion Systems},
author={Wang, Yiyang and Jin, Yiqiao and Cabral, Alex and Hester, Josiah},
booktitle={EMNLP},
year={2026}
}The data, code and model checkpoint are intended and licensed mainly for research.
The source code is released under the
Apache License 2.0
(see THIRD_PARTY_NOTICES.md in the repository for the license scope of figures,
website assets, datasets, and third-party components).
This website is licensed under the Creative Commons Attribution-ShareAlike 4.0 International License.