hello-diana/MASCOT

What it solves

MASCOT addresses the problem of "persona collapse" and "social sycophancy" in multi-agent LLM systems. In these systems, agents often lose their unique identities and regress to generic assistant behavior, or they simply agree with each other unproductively (echo chambers) rather than contributing complementary perspectives.

How it works

MASCOT uses a bi-level optimization strategy to balance individual identity with group dynamics:

1. Persona-Aware Behavioral Alignment: Individual "Speaker" agents are fine-tuned using a Reinforcement-Learning-from-AI-Feedback (RLAIF) pipeline. This involves generating candidate responses, scoring them with a rubric-based LLM judge, training a persona reward model, and optimizing the speaker policy via Group Relative Policy Optimization (GRPO).
2. Collaborative Dialogue Optimization: A "Director" agent is trained to orchestrate the conversation. The Director selects the next speaker and provides natural-language directives (specifying speaker, action, and tone). The Director is optimized using GRPO based on group-level rewards to ensure non-redundant, complementary dialogue.

Who it’s for

Researchers and developers building socio-collaborative AI companions for affective settings (like emotional support) or collaborative professional environments (like workplace meetings).

Highlights

• Bi-level Optimization: Separates the training of individual persona fidelity from the coordination of group interaction.
• Director Orchestration: Uses a dedicated meta-agent to manage turn-taking and tone, preventing redundant contributions.
• Low-Resource Training: Employs LoRA adapters for Speaker and Director policies to keep trainable parameters minimal (0.187%).
• Comprehensive Pipeline: Includes a self-contained workflow for episode generation, reward modeling, GRPO/DPO training, and evaluation.